#include "pipe/p_defines.h"
#include "util/u_inlines.h"
#include "util/u_memory.h"
+#include "util/u_pointer.h"
#include "util/u_format.h"
#include "util/u_dump.h"
#include "os/os_time.h"
LLVMValueRef stencil_refs[2],
LLVMValueRef src,
LLVMValueRef dst_ptr,
- LLVMValueRef facing)
+ LLVMValueRef facing,
+ LLVMValueRef counter)
{
const struct util_format_description *format_desc;
struct lp_type dst_type;
stencil_refs,
src,
dst_ptr,
- facing);
+ facing,
+ counter);
}
LLVMValueRef c2,
LLVMValueRef step0_ptr,
LLVMValueRef step1_ptr,
- LLVMValueRef step2_ptr)
+ LLVMValueRef step2_ptr,
+ LLVMValueRef counter)
{
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef vec_type;
if (early_depth_stencil_test)
generate_depth_stencil(builder, key,
type, &mask,
- stencil_refs, z, depth_ptr, facing);
+ stencil_refs, z, depth_ptr, facing, counter);
lp_build_tgsi_soa(builder, tokens, type, &mask,
consts_ptr, interp->pos, interp->inputs,
outputs, sampler, &shader->info);
+ /* 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]) {
if (!early_depth_stencil_test)
generate_depth_stencil(builder, key,
type, &mask,
- stencil_refs, z, depth_ptr, facing);
+ stencil_refs, z, depth_ptr, facing, counter);
lp_build_mask_end(&mask);
/**
* Generate color blending and color output.
+ * \param rt the render target index (to index blend, colormask state)
+ * \param type the pixel color type
+ * \param context_ptr pointer to the runtime JIT context
+ * \param mask execution mask (active fragment/pixel mask)
+ * \param src colors from the fragment shader
+ * \param dst_ptr the destination color buffer pointer
*/
static void
generate_blend(const struct pipe_blend_state *blend,
+ unsigned rt,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
const_ptr = LLVMBuildBitCast(builder, const_ptr,
LLVMPointerType(vec_type, 0), "");
+ /* load constant blend color and colors from the dest color buffer */
for(chan = 0; chan < 4; ++chan) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
}
- lp_build_blend_soa(builder, blend, type, src, dst, con, res);
+ /* do blend */
+ lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
+ /* store results to color buffer */
for(chan = 0; chan < 4; ++chan) {
- if(blend->rt[0].colormask & (1 << chan)) {
+ if(blend->rt[rt].colormask & (1 << chan)) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
lp_build_name(res[chan], "res.%c", "rgba"[chan]);
res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
}
-/** casting function to avoid compiler warnings */
-static lp_jit_frag_func
-cast_voidptr_to_lp_jit_frag_func(void *p)
-{
- union {
- void *v;
- lp_jit_frag_func f;
- } tmp;
- assert(sizeof(tmp.v) == sizeof(tmp.f));
- tmp.v = p;
- return tmp.f;
-}
-
-
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
- LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
- LLVMTypeRef arg_types[15];
+ LLVMTypeRef arg_types[16];
LLVMTypeRef func_type;
LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
LLVMValueRef context_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
- LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
+ LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr, counter = NULL;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
- LLVMValueRef blend_in_color[NUM_CHANNELS];
LLVMValueRef function;
LLVMValueRef facing;
unsigned num_fs;
*/
fs_elem_type = lp_build_elem_type(fs_type);
- fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
+ arg_types[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
- lp_build_name(color_ptr_ptr, "color_ptr");
+ lp_build_name(color_ptr_ptr, "color_ptr_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(c0, "c0");
lp_build_name(c1, "c1");
lp_build_name(step1_ptr, "step1");
lp_build_name(step2_ptr, "step2");
+ if (key->occlusion_count) {
+ counter = LLVMGetParam(function, 15);
+ lp_build_name(counter, "counter");
+ }
+
/*
* Function body
*/
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp,
- shader->base.tokens,
+ &shader->info,
key->flatshade,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
facing,
do_tri_test,
c0, c1, c2,
- step0_ptr, step1_ptr, step2_ptr);
+ step0_ptr, step1_ptr, step2_ptr, counter);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
+ LLVMValueRef blend_in_color[NUM_CHANNELS];
+ unsigned rt;
/*
* Convert the fs's output color and mask to fit to the blending type.
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
+ /* which blend/colormask state to use */
+ rt = key->blend.independent_blend_enable ? cbuf : 0;
+
/*
* Blending.
*/
generate_blend(&key->blend,
+ rt,
builder,
blend_type,
context_ptr,
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
- LLVMDumpValue(function);
+ lp_debug_dump_value(function);
abort();
}
#endif
if (1)
LLVMRunFunctionPassManager(screen->pass, function);
- if (LP_DEBUG & DEBUG_JIT) {
+ if (gallivm_debug & GALLIVM_DEBUG_IR) {
/* Print the LLVM IR to stderr */
- LLVMDumpValue(function);
+ lp_debug_dump_value(function);
debug_printf("\n");
}
{
void *f = LLVMGetPointerToGlobal(screen->engine, function);
- variant->jit_function[do_tri_test] = cast_voidptr_to_lp_jit_frag_func(f);
+ variant->jit_function[do_tri_test] = (lp_jit_frag_func)pointer_to_func(f);
- if (LP_DEBUG & DEBUG_ASM)
+ if (gallivm_debug & GALLIVM_DEBUG_ASM) {
lp_disassemble(f);
+ }
+ }
+}
+
+
+static void
+dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
+{
+ unsigned i;
+
+ debug_printf("fs variant %p:\n", (void *) key);
+
+ if (key->depth.enabled) {
+ debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
+ debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
+ debug_printf("depth.writemask = %u\n", key->depth.writemask);
+ }
+
+ for (i = 0; i < 2; ++i) {
+ if (key->stencil[i].enabled) {
+ debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
+ debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
+ debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
+ debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
+ debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
+ debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
+ }
+ }
+
+ if (key->alpha.enabled) {
+ debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
+ debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
+ }
+
+ if (key->blend.logicop_enable) {
+ debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
+ }
+ else if (key->blend.rt[0].blend_enable) {
+ debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
+ debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
+ debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
+ debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
+ debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
+ 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);
+ for (i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
+ if (key->sampler[i].format) {
+ debug_printf("sampler[%u] = \n", i);
+ debug_printf(" .format = %s\n",
+ util_format_name(key->sampler[i].format));
+ debug_printf(" .target = %s\n",
+ util_dump_tex_target(key->sampler[i].target, TRUE));
+ debug_printf(" .pot = %u %u %u\n",
+ key->sampler[i].pot_width,
+ key->sampler[i].pot_height,
+ key->sampler[i].pot_depth);
+ debug_printf(" .wrap = %s %s %s\n",
+ util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
+ util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
+ util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
+ debug_printf(" .min_img_filter = %s\n",
+ util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
+ debug_printf(" .min_mip_filter = %s\n",
+ util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
+ debug_printf(" .mag_img_filter = %s\n",
+ util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
+ if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
+ debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
+ debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
+ }
}
}
+
static struct lp_fragment_shader_variant *
generate_variant(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
{
struct lp_fragment_shader_variant *variant;
- if (LP_DEBUG & DEBUG_JIT) {
- unsigned i;
-
+ if (gallivm_debug & GALLIVM_DEBUG_IR) {
tgsi_dump(shader->base.tokens, 0);
- if(key->depth.enabled) {
- debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
- debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
- debug_printf("depth.writemask = %u\n", key->depth.writemask);
- }
- if(key->alpha.enabled) {
- debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
- debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
- }
- if(key->blend.logicop_enable) {
- debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
- }
- else if(key->blend.rt[0].blend_enable) {
- debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
- debug_printf("rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
- debug_printf("rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
- debug_printf("alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
- debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
- debug_printf("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);
- for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
- if(key->sampler[i].format) {
- debug_printf("sampler[%u] = \n", i);
- debug_printf(" .format = %s\n",
- util_format_name(key->sampler[i].format));
- debug_printf(" .target = %s\n",
- util_dump_tex_target(key->sampler[i].target, TRUE));
- debug_printf(" .pot = %u %u %u\n",
- key->sampler[i].pot_width,
- key->sampler[i].pot_height,
- key->sampler[i].pot_depth);
- debug_printf(" .wrap = %s %s %s\n",
- util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
- util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
- util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
- debug_printf(" .min_img_filter = %s\n",
- util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
- debug_printf(" .min_mip_filter = %s\n",
- util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
- debug_printf(" .mag_img_filter = %s\n",
- util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
- if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
- debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
- debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
- }
- }
+ dump_fs_variant_key(key);
}
variant = CALLOC_STRUCT(lp_fragment_shader_variant);
if(!variant)
return NULL;
- variant->shader = shader;
memcpy(&variant->key, key, sizeof *key);
- generate_fragment(lp, shader, variant, 0);
- generate_fragment(lp, shader, variant, 1);
+ generate_fragment(lp, shader, variant, RAST_WHOLE);
+ generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
+
+ /* TODO: most of these can be relaxed, in particular the colormask */
+ variant->opaque =
+ !key->blend.logicop_enable &&
+ !key->blend.rt[0].blend_enable &&
+ key->blend.rt[0].colormask == 0xf &&
+ !key->stencil[0].enabled &&
+ !key->alpha.enabled &&
+ !key->depth.enabled &&
+ !key->scissor &&
+ !shader->info.uses_kill
+ ? TRUE : FALSE;
/* insert new variant into linked list */
variant->next = shader->variants;
}
-void *
+static void *
llvmpipe_create_fs_state(struct pipe_context *pipe,
const struct pipe_shader_state *templ)
{
/* we need to keep a local copy of the tokens */
shader->base.tokens = tgsi_dup_tokens(templ->tokens);
+ if (LP_DEBUG & DEBUG_TGSI) {
+ unsigned attrib;
+ debug_printf("llvmpipe: Create fragment shader %p:\n", (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];
+ debug_printf(" IN[%u].%s%s%s%s\n",
+ attrib,
+ usage_mask & TGSI_WRITEMASK_X ? "x" : "",
+ usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
+ usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
+ usage_mask & TGSI_WRITEMASK_W ? "w" : "");
+ }
+ debug_printf("\n");
+ }
+
return shader;
}
-void
+static void
llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
}
-void
+static void
llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
-void
+static void
llvmpipe_set_constant_buffer(struct pipe_context *pipe,
uint shader, uint index,
struct pipe_resource *constants)
}
+/**
+ * Return the blend factor equivalent to a destination alpha of one.
+ */
+static INLINE unsigned
+force_dst_alpha_one(unsigned factor, boolean alpha)
+{
+ switch(factor) {
+ case PIPE_BLENDFACTOR_DST_ALPHA:
+ return PIPE_BLENDFACTOR_ONE;
+ case PIPE_BLENDFACTOR_INV_DST_ALPHA:
+ return PIPE_BLENDFACTOR_ZERO;
+ case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
+ return PIPE_BLENDFACTOR_ZERO;
+ }
+
+ if (alpha) {
+ switch(factor) {
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ return PIPE_BLENDFACTOR_ONE;
+ case PIPE_BLENDFACTOR_INV_DST_COLOR:
+ return PIPE_BLENDFACTOR_ZERO;
+ }
+ }
+
+ return factor;
+}
+
+
/**
* We need to generate several variants of the fragment pipeline to match
* all the combinations of the contributing state atoms.
key->flatshade = lp->rasterizer->flatshade;
key->scissor = lp->rasterizer->scissor;
+ if (lp->active_query_count) {
+ key->occlusion_count = TRUE;
+ }
if (lp->framebuffer.nr_cbufs) {
memcpy(&key->blend, lp->blend, sizeof key->blend);
key->nr_cbufs = lp->framebuffer.nr_cbufs;
for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
+ struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
const struct util_format_description *format_desc;
unsigned chan;
assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
- key->blend.rt[i].colormask = lp->blend->rt[i].colormask;
+ blend_rt->colormask = lp->blend->rt[i].colormask;
/* mask out color channels not present in the color buffer.
* Should be simple to incorporate per-cbuf writemasks:
enum util_format_swizzle swizzle = format_desc->swizzle[chan];
if(swizzle > UTIL_FORMAT_SWIZZLE_W)
- key->blend.rt[i].colormask &= ~(1 << chan);
+ blend_rt->colormask &= ~(1 << chan);
+ }
+
+ /*
+ * 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.
+ */
+ if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
+ blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE);
+ blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE);
+ blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE);
+ blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE);
}
}
struct lp_fragment_shader *shader = lp->fs;
struct lp_fragment_shader_variant_key key;
struct lp_fragment_shader_variant *variant;
- boolean opaque;
make_variant_key(lp, shader, &key);
LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
}
- shader->current = variant;
+ lp_setup_set_fs_variant(lp->setup, variant);
+}
+
- /* TODO: put this in the variant */
- /* TODO: most of these can be relaxed, in particular the colormask */
- opaque = !key.blend.logicop_enable &&
- !key.blend.rt[0].blend_enable &&
- key.blend.rt[0].colormask == 0xf &&
- !key.stencil[0].enabled &&
- !key.alpha.enabled &&
- !key.depth.enabled &&
- !key.scissor &&
- !shader->info.uses_kill
- ? TRUE : FALSE;
-
- lp_setup_set_fs_functions(lp->setup,
- shader->current->jit_function[RAST_WHOLE],
- shader->current->jit_function[RAST_EDGE_TEST],
- opaque);
+
+void
+llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
+{
+ llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
+ llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;
+ llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
+
+ llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
}