#include "lp_rast.h"
#include "nir/nir_to_tgsi_info.h"
+#include "lp_screen.h"
+#include "compiler/nir/nir_serialize.h"
+#include "util/mesa-sha1.h"
/** Fragment shader number (for debugging) */
static unsigned fs_no = 0;
s_mask = LLVMBuildAnd(builder, s_mask, test, "");
LLVMBuildStore(builder, s_mask, s_mask_ptr);
}
+};
+
+struct lp_build_fs_llvm_iface {
+ struct lp_build_fs_iface base;
+ struct lp_build_interp_soa_context *interp;
+ struct lp_build_for_loop_state *loop_state;
+ LLVMValueRef mask_store;
+};
+
+static LLVMValueRef fs_interp(const struct lp_build_fs_iface *iface,
+ struct lp_build_context *bld,
+ unsigned attrib, unsigned chan,
+ bool centroid, bool sample,
+ LLVMValueRef attrib_indir,
+ LLVMValueRef offsets[2])
+{
+ struct lp_build_fs_llvm_iface *fs_iface = (struct lp_build_fs_llvm_iface *)iface;
+ struct lp_build_interp_soa_context *interp = fs_iface->interp;
+ unsigned loc = TGSI_INTERPOLATE_LOC_CENTER;
+ if (centroid)
+ loc = TGSI_INTERPOLATE_LOC_CENTROID;
+ if (sample)
+ loc = TGSI_INTERPOLATE_LOC_SAMPLE;
+
+ return lp_build_interp_soa(interp, bld->gallivm, fs_iface->loop_state->counter,
+ fs_iface->mask_store,
+ attrib, chan, loc, attrib_indir, offsets);
}
/**
memset(outputs, 0, sizeof outputs);
+ /* Allocate color storage for each fragment sample */
+ LLVMValueRef color_store_size = num_loop;
+ if (key->min_samples > 1)
+ color_store_size = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, key->min_samples), "");
+
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
lp_build_vec_type(gallivm,
type),
- num_loop, "color");
+ color_store_size, "color");
}
}
if (dual_source_blend) {
out_color[1][chan] = lp_build_array_alloca(gallivm,
lp_build_vec_type(gallivm,
type),
- num_loop, "color1");
+ color_store_size, "color1");
}
}
num_loop,
lp_build_const_int32(gallivm, 1));
+ LLVMValueRef sample_mask_in;
if (key->multisample) {
+ sample_mask_in = lp_build_const_int_vec(gallivm, type, 0);
/* create shader execution mask by combining all sample masks. */
for (unsigned s = 0; s < key->coverage_samples; s++) {
LLVMValueRef s_mask_idx = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, s), "");
mask_val = s_mask;
else
mask_val = LLVMBuildOr(builder, s_mask, mask_val, "");
+
+ LLVMValueRef mask_in = LLVMBuildAnd(builder, s_mask, lp_build_const_int_vec(gallivm, type, (1 << s)), "");
+ sample_mask_in = LLVMBuildOr(builder, sample_mask_in, mask_in, "");
}
} else {
+ sample_mask_in = lp_build_const_int_vec(gallivm, type, 1);
mask_ptr = LLVMBuildGEP(builder, mask_store,
&loop_state.counter, 1, "mask_ptr");
mask_val = LLVMBuildLoad(builder, mask_ptr, "");
+
+ LLVMValueRef mask_in = LLVMBuildAnd(builder, mask_val, lp_build_const_int_vec(gallivm, type, 1), "");
+ sample_mask_in = LLVMBuildOr(builder, sample_mask_in, mask_in, "");
}
/* 'mask' will control execution based on quad's pixel alive/killed state */
tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
lp_build_mask_update(&mask, tmp_s_mask_or);
- /* for multisample Z needs to be re interpolated at pixel center */
- lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);
+ if (key->min_samples == 1) {
+ /* for multisample Z needs to be re interpolated at pixel center */
+ lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);
+ lp_build_mask_update(&mask, tmp_s_mask_or);
+ }
}
LLVMValueRef out_sample_mask_storage = NULL;
if (shader->info.base.writes_samplemask) {
out_sample_mask_storage = lp_build_alloca(gallivm, int_vec_type, "write_mask");
+ if (key->min_samples > 1)
+ LLVMBuildStore(builder, LLVMConstNull(int_vec_type), out_sample_mask_storage);
}
+
+ if (key->multisample && key->min_samples > 1) {
+ lp_build_for_loop_begin(&sample_loop_state, gallivm,
+ lp_build_const_int32(gallivm, 0),
+ LLVMIntULT,
+ lp_build_const_int32(gallivm, key->min_samples),
+ lp_build_const_int32(gallivm, 1));
+
+ LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
+ s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
+ s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
+ s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
+ lp_build_mask_force(&mask, s_mask);
+ lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, sample_loop_state.counter);
+ system_values.sample_id = sample_loop_state.counter;
+ } else
+ system_values.sample_id = lp_build_const_int32(gallivm, 0);
+
+ system_values.sample_mask_in = sample_mask_in;
system_values.sample_pos = sample_pos_array;
- lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, mask_store, NULL);
+ lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, mask_store, sample_loop_state.counter);
+
+ struct lp_build_fs_llvm_iface fs_iface = {
+ .base.interp_fn = fs_interp,
+ .interp = interp,
+ .loop_state = &loop_state,
+ .mask_store = mask_store,
+ };
struct lp_build_tgsi_params params;
memset(¶ms, 0, sizeof(params));
params.type = type;
params.mask = &mask;
+ params.fs_iface = &fs_iface.base;
params.consts_ptr = consts_ptr;
params.const_sizes_ptr = num_consts_ptr;
params.system_values = &system_values;
assert(smaski >= 0);
output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
- /*
- * Pixel is alive according to the first sample in the mask.
- */
output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");
- if (!key->multisample) {
- output_smask = lp_build_and(&smask_bld, output_smask, smask_bld.one);
- output_smask = lp_build_cmp(&smask_bld, PIPE_FUNC_NOTEQUAL, output_smask, smask_bld.zero);
- lp_build_mask_update(&mask, output_smask);
+
+ if (key->min_samples > 1) {
+ /* only the bit corresponding to this sample is to be used. */
+ LLVMValueRef tmp_mask = LLVMBuildLoad(builder, out_sample_mask_storage, "tmp_mask");
+ LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
+ LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, lp_build_broadcast(gallivm, int_vec_type, out_smask_idx), "");
+ output_smask = LLVMBuildOr(builder, tmp_mask, smask_bit, "");
}
+
LLVMBuildStore(builder, output_smask, out_sample_mask_storage);
}
+ /* Color write - per fragment sample */
+ for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
+ {
+ unsigned cbuf = shader->info.base.output_semantic_index[attrib];
+ if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
+ ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
+ {
+ for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
+ if(outputs[attrib][chan]) {
+ /* XXX: just initialize outputs to point at colors[] and
+ * skip this.
+ */
+ LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
+ LLVMValueRef color_ptr;
+ LLVMValueRef color_idx = loop_state.counter;
+ if (key->min_samples > 1)
+ color_idx = LLVMBuildAdd(builder, color_idx,
+ LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
+ color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
+ &color_idx, 1, "");
+ lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
+ LLVMBuildStore(builder, out, color_ptr);
+ }
+ }
+ }
+ }
+
+ if (key->multisample && key->min_samples > 1) {
+ LLVMBuildStore(builder, lp_build_mask_value(&mask), s_mask_ptr);
+ lp_build_for_loop_end(&sample_loop_state);
+ }
+
if (key->multisample) {
/* execute depth test for each sample */
lp_build_for_loop_begin(&sample_loop_state, gallivm,
/* combine the execution mask post fragment shader with the coverage mask. */
s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
- s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
+ if (key->min_samples == 1)
+ s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
/* if the shader writes sample mask use that */
if (shader->info.base.writes_samplemask) {
z_value, s_value);
}
- if (key->multisample) {
- /* store the sample mask for this loop */
- LLVMBuildStore(builder, s_mask, s_mask_ptr);
- lp_build_for_loop_end(&sample_loop_state);
- }
-
- /* Color write */
- for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
- {
- unsigned cbuf = shader->info.base.output_semantic_index[attrib];
- if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
- ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
- {
- for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
- if(outputs[attrib][chan]) {
- /* XXX: just initialize outputs to point at colors[] and
- * skip this.
- */
- LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
- LLVMValueRef color_ptr;
- color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
- &loop_state.counter, 1, "");
- lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
- LLVMBuildStore(builder, out, color_ptr);
- }
- }
- }
- }
-
if (key->occlusion_count) {
LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
lp_build_name(counter, "counter");
+
lp_build_occlusion_count(gallivm, type,
- lp_build_mask_value(&mask), counter);
+ key->multisample ? s_mask : lp_build_mask_value(&mask), counter);
+ }
+
+ if (key->multisample) {
+ /* store the sample mask for this loop */
+ LLVMBuildStore(builder, s_mask, s_mask_ptr);
+ lp_build_for_loop_end(&sample_loop_state);
}
mask_val = lp_build_mask_end(&mask);
struct lp_build_image_soa *image;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
- LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
+ LLVMValueRef fs_out_color[LP_MAX_SAMPLES][PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
LLVMValueRef function;
LLVMValueRef facing;
unsigned num_fs;
blend_vec_type = lp_build_vec_type(gallivm, blend_type);
- snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
- shader->no, variant->no, partial_mask ? "partial" : "whole");
+ snprintf(func_name, sizeof(func_name), "fs_variant_%s",
+ partial_mask ? "partial" : "whole");
arg_types[0] = variant->jit_context_ptr_type; /* context */
arg_types[1] = int32_type; /* x */
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
+ if (variant->gallivm->cache->data_size)
+ return;
+
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
}
/* code generated texture sampling */
- sampler = lp_llvm_sampler_soa_create(key->samplers);
+ sampler = lp_llvm_sampler_soa_create(key->samplers, key->nr_samplers);
image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key));
num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
num_loop_samp, "mask_store");
LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
- LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, flt_type, "");
+ LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, LLVMArrayType(flt_type, key->coverage_samples * 2), "");
LLVMValueRef sample_pos_array;
if (key->multisample && key->coverage_samples == 4) {
thread_data_ptr);
for (i = 0; i < num_fs; i++) {
- LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
LLVMValueRef ptr;
for (unsigned s = 0; s < key->coverage_samples; s++) {
int idx = (i + (s * num_fs));
fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
}
- /* This is fucked up need to reorganize things */
- for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
- for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
- ptr = LLVMBuildGEP(builder,
- color_store[cbuf * !cbuf0_write_all][chan],
- &indexi, 1, "");
- fs_out_color[cbuf][chan][i] = ptr;
+ for (unsigned s = 0; s < key->min_samples; s++) {
+ /* This is fucked up need to reorganize things */
+ int idx = s * num_fs + i;
+ LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
+ for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
+ for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
+ ptr = LLVMBuildGEP(builder,
+ color_store[cbuf * !cbuf0_write_all][chan],
+ &sindexi, 1, "");
+ fs_out_color[s][cbuf][chan][i] = ptr;
+ }
}
- }
- if (dual_source_blend) {
- /* only support one dual source blend target hence always use output 1 */
- for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
- ptr = LLVMBuildGEP(builder,
- color_store[1][chan],
- &indexi, 1, "");
- fs_out_color[1][chan][i] = ptr;
+ if (dual_source_blend) {
+ /* only support one dual source blend target hence always use output 1 */
+ for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
+ ptr = LLVMBuildGEP(builder,
+ color_store[1][chan],
+ &sindexi, 1, "");
+ fs_out_color[s][1][chan][i] = ptr;
+ }
}
}
}
for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {
unsigned mask_idx = num_fs * (key->multisample ? s : 0);
+ unsigned out_idx = key->min_samples == 1 ? 0 : s;
LLVMValueRef out_ptr = color_ptr;;
if (key->multisample) {
generate_unswizzled_blend(gallivm, cbuf, variant,
key->cbuf_format[cbuf],
- num_fs, fs_type, &fs_mask[mask_idx], fs_out_color,
+ num_fs, fs_type, &fs_mask[mask_idx], fs_out_color[out_idx],
context_ptr, out_ptr, stride,
partial_mask, do_branch);
}
if (key->multisample) {
debug_printf("multisample = 1\n");
debug_printf("coverage samples = %d\n", key->coverage_samples);
+ debug_printf("min samples = %d\n", key->min_samples);
}
for (i = 0; i < key->nr_cbufs; ++i) {
debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
debug_printf("\n");
}
+static void
+lp_fs_get_ir_cache_key(struct lp_fragment_shader_variant *variant,
+ unsigned char ir_sha1_cache_key[20])
+{
+ struct blob blob = { 0 };
+ unsigned ir_size;
+ void *ir_binary;
+
+ blob_init(&blob);
+ nir_serialize(&blob, variant->shader->base.ir.nir, true);
+ ir_binary = blob.data;
+ ir_size = blob.size;
+
+ struct mesa_sha1 ctx;
+ _mesa_sha1_init(&ctx);
+ _mesa_sha1_update(&ctx, &variant->key, variant->shader->variant_key_size);
+ _mesa_sha1_update(&ctx, ir_binary, ir_size);
+ _mesa_sha1_final(&ctx, ir_sha1_cache_key);
+
+ blob_finish(&blob);
+}
/**
* Generate a new fragment shader variant from the shader code and
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key)
{
+ struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
struct lp_fragment_shader_variant *variant;
const struct util_format_description *cbuf0_format_desc = NULL;
boolean fullcolormask;
char module_name[64];
-
+ unsigned char ir_sha1_cache_key[20];
+ struct lp_cached_code cached = { 0 };
+ bool needs_caching = false;
variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
if (!variant)
return NULL;
snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
shader->no, shader->variants_created);
- variant->gallivm = gallivm_create(module_name, lp->context);
+ variant->shader = shader;
+ memcpy(&variant->key, key, shader->variant_key_size);
+
+ if (shader->base.ir.nir) {
+ lp_fs_get_ir_cache_key(variant, ir_sha1_cache_key);
+
+ lp_disk_cache_find_shader(screen, &cached, ir_sha1_cache_key);
+ if (!cached.data_size)
+ needs_caching = true;
+ }
+ variant->gallivm = gallivm_create(module_name, lp->context, &cached);
if (!variant->gallivm) {
FREE(variant);
return NULL;
}
- variant->shader = shader;
variant->list_item_global.base = variant;
variant->list_item_local.base = variant;
variant->no = shader->variants_created++;
- memcpy(&variant->key, key, shader->variant_key_size);
+
/*
* Determine whether we are touching all channels in the color buffer.
fullcolormask &&
!key->stencil[0].enabled &&
!key->alpha.enabled &&
+ !key->multisample &&
!key->blend.alpha_to_coverage &&
!key->depth.enabled &&
!shader->info.base.uses_kill &&
variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
}
+ if (needs_caching) {
+ lp_disk_cache_insert_shader(screen, &cached, ir_sha1_cache_key);
+ }
+
gallivm_free_ir(variant->gallivm);
return variant;
}
key->coverage_samples = 1;
- if (key->multisample)
+ key->min_samples = 1;
+ if (key->multisample) {
key->coverage_samples = util_framebuffer_get_num_samples(&lp->framebuffer);
+ key->min_samples = lp->min_samples == 1 ? 1 : key->coverage_samples;
+ }
key->nr_cbufs = lp->framebuffer.nr_cbufs;
if (!key->blend.independent_blend_enable) {
projects / mesa.git / blobdiff