#include "util/u_inlines.h"
#include "util/u_memory.h"
#include "util/u_pointer.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_dump.h"
#include "util/u_string.h"
#include "util/simple_list.h"
#include "util/u_dual_blend.h"
-#include "os/os_time.h"
+#include "util/os_time.h"
#include "pipe/p_shader_tokens.h"
#include "draw/draw_context.h"
#include "tgsi/tgsi_dump.h"
#include "gallivm/lp_bld_intr.h"
#include "gallivm/lp_bld_logic.h"
#include "gallivm/lp_bld_tgsi.h"
+#include "gallivm/lp_bld_nir.h"
#include "gallivm/lp_bld_swizzle.h"
#include "gallivm/lp_bld_flow.h"
#include "gallivm/lp_bld_debug.h"
#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_bitarit.h"
#include "gallivm/lp_bld_pack.h"
#include "gallivm/lp_bld_format.h"
#include "gallivm/lp_bld_quad.h"
#include "lp_flush.h"
#include "lp_state_fs.h"
#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;
generate_quad_mask(struct gallivm_state *gallivm,
struct lp_type fs_type,
unsigned first_quad,
- LLVMValueRef mask_input) /* int32 */
+ unsigned sample,
+ LLVMValueRef mask_input) /* int64 */
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type mask_type;
shift = 0;
}
+ mask_input = LLVMBuildLShr(builder, mask_input, lp_build_const_int64(gallivm, 16 * sample), "");
+ mask_input = LLVMBuildTrunc(builder, mask_input,
+ i32t, "");
+ mask_input = LLVMBuildAnd(builder, mask_input, lp_build_const_int32(gallivm, 0xffff), "");
+
mask_input = LLVMBuildLShr(builder,
mask_input,
LLVMConstInt(i32t, shift, 0),
return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
}
+static void
+lp_build_sample_alpha_to_coverage(struct gallivm_state *gallivm,
+ struct lp_type type,
+ unsigned coverage_samples,
+ LLVMValueRef num_loop,
+ LLVMValueRef loop_counter,
+ LLVMValueRef coverage_mask_store,
+ LLVMValueRef alpha)
+{
+ struct lp_build_context bld;
+ LLVMBuilderRef builder = gallivm->builder;
+ float step = 1.0 / coverage_samples;
+
+ lp_build_context_init(&bld, gallivm, type);
+ for (unsigned s = 0; s < coverage_samples; s++) {
+ LLVMValueRef alpha_ref_value = lp_build_const_vec(gallivm, type, step * s);
+ LLVMValueRef test = lp_build_cmp(&bld, PIPE_FUNC_GREATER, alpha, alpha_ref_value);
+
+ LLVMValueRef s_mask_idx = LLVMBuildMul(builder, lp_build_const_int32(gallivm, s), num_loop, "");
+ s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_counter, "");
+ LLVMValueRef s_mask_ptr = LLVMBuildGEP(builder, coverage_mask_store, &s_mask_idx, 1, "");
+ LLVMValueRef s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
+ 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);
+}
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
+ LLVMValueRef sample_pos_array,
LLVMValueRef num_loop,
struct lp_build_interp_soa_context *interp,
- struct lp_build_sampler_soa *sampler,
+ const struct lp_build_sampler_soa *sampler,
+ const struct lp_build_image_soa *image,
LLVMValueRef mask_store,
LLVMValueRef (*out_color)[4],
- LLVMValueRef depth_ptr,
+ LLVMValueRef depth_base_ptr,
LLVMValueRef depth_stride,
+ LLVMValueRef depth_sample_stride,
LLVMValueRef facing,
LLVMValueRef thread_data_ptr)
{
const struct tgsi_token *tokens = shader->base.tokens;
struct lp_type int_type = lp_int_type(type);
LLVMTypeRef vec_type, int_vec_type;
- LLVMValueRef mask_ptr, mask_val;
+ LLVMValueRef mask_ptr = NULL, mask_val = NULL;
LLVMValueRef consts_ptr, num_consts_ptr;
+ LLVMValueRef ssbo_ptr, num_ssbo_ptr;
LLVMValueRef z;
LLVMValueRef z_value, s_value;
LLVMValueRef z_fb, s_fb;
+ LLVMValueRef depth_ptr;
LLVMValueRef stencil_refs[2];
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
- struct lp_build_for_loop_state loop_state;
+ LLVMValueRef zs_samples = lp_build_const_int32(gallivm, key->zsbuf_nr_samples);
+ struct lp_build_for_loop_state loop_state, sample_loop_state;
struct lp_build_mask_context mask;
/*
* TODO: figure out if simple_shader optimization is really worthwile to
memset(&system_values, 0, sizeof(system_values));
+ /* truncate then sign extend. */
+ system_values.front_facing = LLVMBuildTrunc(gallivm->builder, facing, LLVMInt1TypeInContext(gallivm->context), "");
+ system_values.front_facing = LLVMBuildSExt(gallivm->builder, system_values.front_facing, LLVMInt32TypeInContext(gallivm->context), "");
+
if (key->depth.enabled ||
key->stencil[0].enabled) {
zs_format_desc = util_format_description(key->zsbuf_format);
assert(zs_format_desc);
- if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
- if (key->alpha.enabled ||
+ if (shader->info.base.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL])
+ depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
+ else if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
+ if (shader->info.base.writes_memory)
+ depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
+ else if (key->alpha.enabled ||
key->blend.alpha_to_coverage ||
- shader->info.base.uses_kill) {
+ shader->info.base.uses_kill ||
+ shader->info.base.writes_samplemask) {
/* 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
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),
- LLVMIntULT,
- num_loop,
- lp_build_const_int32(gallivm, 1));
-
- mask_ptr = LLVMBuildGEP(builder, mask_store,
- &loop_state.counter, 1, "mask_ptr");
- mask_val = LLVMBuildLoad(builder, mask_ptr, "");
+ ssbo_ptr = lp_jit_context_ssbos(gallivm, context_ptr);
+ num_ssbo_ptr = lp_jit_context_num_ssbos(gallivm, context_ptr);
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");
}
}
+ lp_build_for_loop_begin(&loop_state, gallivm,
+ lp_build_const_int32(gallivm, 0),
+ LLVMIntULT,
+ 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), "");
+ s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
+ LLVMValueRef s_mask = lp_build_pointer_get(builder, mask_store, s_mask_idx);
+ if (s == 0)
+ 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 */
lp_build_mask_begin(&mask, gallivm, type, mask_val);
if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
lp_build_mask_check(&mask);
- lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
+ /* Create storage for recombining sample masks after early Z pass. */
+ LLVMValueRef s_mask_or = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, type), "cov_mask_early_depth");
+ LLVMBuildStore(builder, LLVMConstNull(lp_build_int_vec_type(gallivm, type)), s_mask_or);
+
+ LLVMValueRef s_mask = NULL, s_mask_ptr = NULL;
+ LLVMValueRef z_sample_value_store = NULL, s_sample_value_store = NULL;
+ LLVMValueRef z_fb_store = NULL, s_fb_store = NULL;
+ LLVMTypeRef z_type = NULL, z_fb_type = NULL;
+
+ /* Run early depth once per sample */
+ if (key->multisample) {
+
+ if (zs_format_desc) {
+ struct lp_type zs_type = lp_depth_type(zs_format_desc, type.length);
+ struct lp_type z_type = zs_type;
+ struct lp_type s_type = zs_type;
+ if (zs_format_desc->block.bits < type.width)
+ z_type.width = type.width;
+ else if (zs_format_desc->block.bits > 32) {
+ z_type.width = z_type.width / 2;
+ s_type.width = s_type.width / 2;
+ s_type.floating = 0;
+ }
+ z_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
+ zs_samples, "z_sample_store");
+ s_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
+ zs_samples, "s_sample_store");
+ z_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, z_type),
+ zs_samples, "z_fb_store");
+ s_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, s_type),
+ zs_samples, "s_fb_store");
+ }
+ lp_build_for_loop_begin(&sample_loop_state, gallivm,
+ lp_build_const_int32(gallivm, 0),
+ LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_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, "");
+ s_mask = LLVMBuildAnd(builder, s_mask, mask_val, "");
+ }
+
+
+ /* for multisample Z needs to be interpolated at sample points for testing. */
+ lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);
z = interp->pos[2];
+ depth_ptr = depth_base_ptr;
+ if (key->multisample) {
+ LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
+ depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
+ }
+
if (depth_mode & EARLY_DEPTH_TEST) {
/*
* Clamp according to ARB_depth_clamp semantics.
key->stencil,
type,
zs_format_desc,
- &mask,
+ key->multisample ? NULL : &mask,
+ &s_mask,
stencil_refs,
z, z_fb, s_fb,
facing,
&z_value, &s_value,
- !simple_shader);
+ !simple_shader && !key->multisample);
if (depth_mode & EARLY_DEPTH_WRITE) {
lp_build_depth_stencil_write_swizzled(gallivm, type,
* stencil test otherwise new stencil values may not get written if all
* fragments got killed by depth/stencil test.
*/
- if (!simple_shader && key->stencil[0].enabled)
+ if (!simple_shader && key->stencil[0].enabled && !key->multisample)
lp_build_mask_check(&mask);
+
+ if (key->multisample) {
+ z_fb_type = LLVMTypeOf(z_fb);
+ z_type = LLVMTypeOf(z_value);
+ lp_build_pointer_set(builder, z_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, z_value, lp_build_int_vec_type(gallivm, type), ""));
+ lp_build_pointer_set(builder, s_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, s_value, lp_build_int_vec_type(gallivm, type), ""));
+ lp_build_pointer_set(builder, z_fb_store, sample_loop_state.counter, z_fb);
+ lp_build_pointer_set(builder, s_fb_store, sample_loop_state.counter, s_fb);
+ }
}
- lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
+ if (key->multisample) {
+ /*
+ * Store the post-early Z coverage mask.
+ * Recombine the resulting coverage masks post early Z into the fragment
+ * shader execution mask.
+ */
+ LLVMValueRef tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
+ tmp_s_mask_or = LLVMBuildOr(builder, tmp_s_mask_or, s_mask, "");
+ LLVMBuildStore(builder, tmp_s_mask_or, s_mask_or);
+
+ LLVMBuildStore(builder, s_mask, s_mask_ptr);
+
+ lp_build_for_loop_end(&sample_loop_state);
+
+ /* recombined all the coverage masks in the shader exec mask. */
+ tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
+ lp_build_mask_update(&mask, tmp_s_mask_or);
+
+ 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, 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;
+ params.inputs = interp->inputs;
+ params.context_ptr = context_ptr;
+ params.thread_data_ptr = thread_data_ptr;
+ params.sampler = sampler;
+ params.info = &shader->info.base;
+ params.ssbo_ptr = ssbo_ptr;
+ params.ssbo_sizes_ptr = num_ssbo_ptr;
+ params.image = image;
/* Build the actual shader */
- lp_build_tgsi_soa(gallivm, tokens, type, &mask,
- consts_ptr, num_consts_ptr, &system_values,
- interp->inputs,
- outputs, context_ptr, thread_data_ptr,
- sampler, &shader->info.base, NULL);
+ if (shader->base.type == PIPE_SHADER_IR_TGSI)
+ lp_build_tgsi_soa(gallivm, tokens, ¶ms,
+ outputs);
+ else
+ lp_build_nir_soa(gallivm, shader->base.ir.nir, ¶ms,
+ outputs);
/* Alpha test */
if (key->alpha.enabled) {
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);
+ if (!key->multisample) {
+ lp_build_alpha_to_coverage(gallivm, type,
+ &mask, alpha,
+ (depth_mode & LATE_DEPTH_TEST) != 0);
+ } else {
+ lp_build_sample_alpha_to_coverage(gallivm, type, key->coverage_samples, num_loop,
+ loop_state.counter,
+ mask_store, alpha);
+ }
+ }
+ }
+ if (key->blend.alpha_to_one && key->multisample) {
+ 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)))
+ if (outputs[cbuf][3]) {
+ LLVMBuildStore(builder, lp_build_const_vec(gallivm, type, 1.0), outputs[cbuf][3]);
+ }
+ }
+ }
+ if (shader->info.base.writes_samplemask) {
+ LLVMValueRef output_smask = NULL;
+ int smaski = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_SAMPLEMASK,
+ 0);
+ struct lp_build_context smask_bld;
+ lp_build_context_init(&smask_bld, gallivm, int_type);
+
+ assert(smaski >= 0);
+ output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
+ output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");
+
+ 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,
+ lp_build_const_int32(gallivm, 0),
+ LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
+ lp_build_const_int32(gallivm, 1));
+
+ /* load the per-sample coverage mask */
+ 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, "");
+
+ /* combine the execution mask post fragment shader with the coverage mask. */
+ s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
+ 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) {
+ LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
+ out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);
+ LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");
+ LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");
+ LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");
+ smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");
+
+ s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
+ }
+ }
+
+ depth_ptr = depth_base_ptr;
+ if (key->multisample) {
+ LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
+ depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
+ }
+
/* Late Z test */
if (depth_mode & LATE_DEPTH_TEST) {
int pos0 = find_output_by_semantic(&shader->info.base,
key->stencil,
type,
zs_format_desc,
- &mask,
+ key->multisample ? NULL : &mask,
+ &s_mask,
stencil_refs,
z, z_fb, s_fb,
facing,
* depth value, update from zs_value with the new mask value and
* write that out.
*/
+ if (key->multisample) {
+ z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;
+ s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);
+ z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");
+ s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);
+ }
lp_build_depth_stencil_write_swizzled(gallivm, type,
zs_format_desc, key->resource_1d,
- &mask, z_fb, s_fb, loop_state.counter,
+ key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,
depth_ptr, depth_stride,
z_value, s_value);
}
-
- /* 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);
- LLVMBuildStore(builder, mask_val, mask_ptr);
+ if (!key->multisample)
+ LLVMBuildStore(builder, mask_val, mask_ptr);
lp_build_for_loop_end(&loop_state);
}
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
unsigned sa = src_fmt->channel[j].shift;
-#ifdef PIPE_ARCH_LITTLE_ENDIAN
+#if UTIL_ARCH_LITTLE_ENDIAN
unsigned from_lsb = j;
#else
unsigned from_lsb = src_fmt->nr_channels - j - 1;
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
unsigned sa = src_fmt->channel[j].shift;
-#ifdef PIPE_ARCH_LITTLE_ENDIAN
+ unsigned sz_a = src_fmt->channel[j].size;
+#if UTIL_ARCH_LITTLE_ENDIAN
unsigned from_lsb = j;
#else
unsigned from_lsb = src_fmt->nr_channels - j - 1;
if (src_type.norm) {
chans[j] = scale_bits(gallivm, blend_type.width,
src_fmt->channel[j].size, chans[j], src_type);
+ } else if (!src_type.floating && sz_a < blend_type.width) {
+ LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);
+ LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans[j], mask_val, "");
+ chans[j] = LLVMBuildSelect(builder, mask, mask_val, chans[j], "");
}
/* Insert bits */
if (dst_count > src_count) {
if ((dst_type.width == 8 || dst_type.width == 16) &&
- util_is_power_of_two(dst_type.length) &&
+ util_is_power_of_two_or_zero(dst_type.length) &&
dst_type.length * dst_type.width < 128) {
/*
* Never try to load values as 4xi8 which we will then
unsigned partial_mask)
{
struct gallivm_state *gallivm = variant->gallivm;
- const struct lp_fragment_shader_variant_key *key = &variant->key;
+ struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
char func_name[64];
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef blend_vec_type;
- LLVMTypeRef arg_types[13];
+ LLVMTypeRef arg_types[15];
LLVMTypeRef func_type;
LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef stride_ptr;
+ LLVMValueRef color_sample_stride_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef depth_stride;
+ LLVMValueRef depth_sample_stride;
LLVMValueRef mask_input;
LLVMValueRef thread_data_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
struct lp_build_sampler_soa *sampler;
+ struct lp_build_image_soa *image;
struct lp_build_interp_soa_context interp;
- LLVMValueRef fs_mask[16 / 4];
- LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
+ LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
+ 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);
- util_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 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
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[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0); /* color */
arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
- arg_types[9] = int32_type; /* mask_input */
+ arg_types[9] = LLVMInt64TypeInContext(gallivm->context); /* mask_input */
arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */
arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */
arg_types[12] = int32_type; /* depth_stride */
+ arg_types[13] = LLVMPointerType(int32_type, 0); /* color sample strides */
+ arg_types[14] = int32_type; /* depth sample stride */
func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
arg_types, ARRAY_SIZE(arg_types), 0);
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);
thread_data_ptr = LLVMGetParam(function, 10);
stride_ptr = LLVMGetParam(function, 11);
depth_stride = LLVMGetParam(function, 12);
+ color_sample_stride_ptr = LLVMGetParam(function, 13);
+ depth_sample_stride = LLVMGetParam(function, 14);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(thread_data_ptr, "thread_data");
lp_build_name(stride_ptr, "stride_ptr");
lp_build_name(depth_stride, "depth_stride");
+ lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");
+ lp_build_name(depth_sample_stride, "depth_sample_stride");
/*
* Function body
assert(builder);
LLVMPositionBuilderAtEnd(builder, block);
+ /*
+ * Must not count ps invocations if there's a null shader.
+ * (It would be ok to count with null shader if there's d/s tests,
+ * but only if there's d/s buffers too, which is different
+ * to implicit rasterization disable which must not depend
+ * on the d/s buffers.)
+ * Could use popcount on mask, but pixel accuracy is not required.
+ * Could disable if there's no stats query, but maybe not worth it.
+ */
+ if (shader->info.base.num_instructions > 1) {
+ LLVMValueRef invocs, val;
+ invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
+ val = LLVMBuildLoad(builder, invocs, "");
+ val = LLVMBuildAdd(builder, val,
+ LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
+ "invoc_count");
+ LLVMBuildStore(builder, val, invocs);
+ }
+
/* code generated texture sampling */
- sampler = lp_llvm_sampler_soa_create(key->state);
+ 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 */
/* for 1d resources only run "upper half" of stamp */
{
LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
+ LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);
LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
- num_loop, "mask_store");
+ num_loop_samp, "mask_store");
+
+ LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
+ 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) {
+ LLVMValueRef sample_pos_arr[8];
+ for (unsigned i = 0; i < 4; i++) {
+ sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);
+ sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);
+ }
+ sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);
+ } else {
+ LLVMValueRef sample_pos_arr[2];
+ sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);
+ sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);
+ sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);
+ }
+ LLVMSetInitializer(glob_sample_pos, sample_pos_array);
+
LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
boolean pixel_center_integer =
shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
shader->info.base.num_inputs,
inputs,
pixel_center_integer,
+ key->coverage_samples, glob_sample_pos,
+ num_loop,
key->depth_clamp,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
for (i = 0; i < num_fs; i++) {
- LLVMValueRef mask;
- LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
- LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
- &indexi, 1, "mask_ptr");
-
- if (partial_mask) {
- mask = generate_quad_mask(gallivm, fs_type,
- i*fs_type.length/4, mask_input);
- }
- else {
- mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
+ if (key->multisample) {
+ LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");
+
+ /*
+ * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
+ * store to the per sample mask storage. Or all of them together to generate
+ * the fragment shader mask. (sample shading TODO).
+ * Take the incoming state coverage mask into account.
+ */
+ for (unsigned s = 0; s < key->coverage_samples; s++) {
+ LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));
+ LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,
+ &sindexi, 1, "sample_mask_ptr");
+ LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,
+ i*fs_type.length/4, s, mask_input);
+
+ LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");
+ LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");
+ smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");
+ smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);
+
+ s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
+ LLVMBuildStore(builder, s_mask, sample_mask_ptr);
+ }
+ } else {
+ LLVMValueRef mask;
+ LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
+ LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
+ &indexi, 1, "mask_ptr");
+
+ if (partial_mask) {
+ mask = generate_quad_mask(gallivm, fs_type,
+ i*fs_type.length/4, 0, mask_input);
+ }
+ else {
+ mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
+ }
+ LLVMBuildStore(builder, mask, mask_ptr);
}
- LLVMBuildStore(builder, mask, mask_ptr);
}
generate_fs_loop(gallivm,
builder,
fs_type,
context_ptr,
+ glob_sample_pos,
num_loop,
&interp,
sampler,
+ image,
mask_store, /* output */
color_store,
depth_ptr,
depth_stride,
+ depth_sample_stride,
facing,
thread_data_ptr);
for (i = 0; i < num_fs; i++) {
- LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
- LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
- &indexi, 1, "");
- fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
- /* 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;
- }
+ LLVMValueRef ptr;
+ for (unsigned s = 0; s < key->coverage_samples; s++) {
+ int idx = (i + (s * num_fs));
+ LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
+ ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");
+
+ fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
}
- 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;
+
+ 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],
+ &sindexi, 1, "");
+ fs_out_color[s][1][chan][i] = ptr;
+ }
}
}
}
}
sampler->destroy(sampler);
-
+ image->destroy(image);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
LLVMValueRef color_ptr;
LLVMValueRef stride;
+ LLVMValueRef sample_stride = NULL;
LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
boolean do_branch = ((key->depth.enabled
&index, 1, ""),
"");
- lp_build_name(color_ptr, "color_ptr%d", cbuf);
-
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);
+ if (key->multisample)
+ sample_stride = LLVMBuildLoad(builder,
+ LLVMBuildGEP(builder, color_sample_stride_ptr,
+ &index, 1, ""), "");
+
+ 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) {
+ LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");
+ out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");
+ }
+ out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");
+
+ lp_build_name(out_ptr, "color_ptr%d", cbuf);
+
+ generate_unswizzled_blend(gallivm, cbuf, variant,
+ key->cbuf_format[cbuf],
+ num_fs, fs_type, &fs_mask[mask_idx], fs_out_color[out_idx],
+ context_ptr, out_ptr, stride,
+ partial_mask, do_branch);
+ }
}
}
static void
-dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
+dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
{
unsigned i;
if (key->flatshade) {
debug_printf("flatshade = 1\n");
}
+ 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("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);
}
if (key->depth.enabled || key->stencil[0].enabled) {
debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
+ debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);
}
if (key->depth.enabled) {
debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
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;
+ const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
debug_printf("sampler[%u] = \n", i);
debug_printf(" .wrap = %s %s %s\n",
util_str_tex_wrap(sampler->wrap_s, TRUE),
debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
}
for (i = 0; i < key->nr_sampler_views; ++i) {
- const struct lp_static_texture_state *texture = &key->state[i].texture_state;
+ const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
debug_printf("texture[%u] = \n", i);
debug_printf(" .format = %s\n",
util_format_name(texture->format));
texture->pot_height,
texture->pot_depth);
}
+ struct lp_image_static_state *images = lp_fs_variant_key_images(key);
+ for (i = 0; i < key->nr_images; ++i) {
+ const struct lp_static_texture_state *image = &images[i].image_state;
+ debug_printf("image[%u] = \n", i);
+ debug_printf(" .format = %s\n",
+ util_format_name(image->format));
+ debug_printf(" .target = %s\n",
+ util_str_tex_target(image->target, TRUE));
+ debug_printf(" .level_zero_only = %u\n",
+ image->level_zero_only);
+ debug_printf(" .pot = %u %u %u\n",
+ image->pot_width,
+ image->pot_height,
+ image->pot_depth);
+ }
}
void
-lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
+lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
{
- debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
+ debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
variant->shader->no, variant->no);
- tgsi_dump(variant->shader->base.tokens, 0);
+ if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
+ tgsi_dump(variant->shader->base.tokens, 0);
+ else
+ nir_print_shader(variant->shader->base.ir.nir, stderr);
dump_fs_variant_key(&variant->key);
debug_printf("variant->opaque = %u\n", variant->opaque);
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;
+ const struct util_format_description *cbuf0_format_desc = NULL;
boolean fullcolormask;
char module_name[64];
-
- variant = CALLOC_STRUCT(lp_fragment_shader_variant);
+ 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;
- util_snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
- shader->no, shader->variants_created);
+ memset(variant, 0, sizeof(*variant));
+ snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
+ shader->no, shader->variants_created);
+
+ variant->shader = shader;
+ memcpy(&variant->key, key, shader->variant_key_size);
- variant->gallivm = gallivm_create(module_name, lp->context);
+ 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
+ !shader->info.base.uses_kill &&
+ !shader->info.base.writes_samplemask
? 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);
}
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;
struct lp_fragment_shader *shader;
int nr_samplers;
int nr_sampler_views;
+ int nr_images;
int i;
shader = CALLOC_STRUCT(lp_fragment_shader);
shader->no = fs_no++;
make_empty_list(&shader->variants);
- /* get/save the summary info for this shader */
- lp_build_tgsi_info(templ->tokens, &shader->info);
+ shader->base.type = templ->type;
+ if (templ->type == PIPE_SHADER_IR_TGSI) {
+ /* get/save the summary info for this shader */
+ 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);
+ /* we need to keep a local copy of the tokens */
+ shader->base.tokens = tgsi_dup_tokens(templ->tokens);
+ } else {
+ shader->base.ir.nir = templ->ir.nir;
+ nir_tgsi_scan_shader(templ->ir.nir, &shader->info.base, true);
+ }
shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
if (shader->draw_data == NULL) {
nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
-
- shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
- state[MAX2(nr_samplers, nr_sampler_views)]);
+ nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
+ shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
for (i = 0; i < shader->info.base.num_inputs; i++) {
shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
+ shader->inputs[i].location = shader->info.base.input_interpolate_loc[i];
switch (shader->info.base.input_interpolate[i]) {
case TGSI_INTERPOLATE_CONSTANT:
llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
-
- if (llvmpipe->fs == fs)
+ struct lp_fragment_shader *lp_fs = (struct lp_fragment_shader *)fs;
+ if (llvmpipe->fs == lp_fs)
return;
- llvmpipe->fs = (struct lp_fragment_shader *) fs;
-
draw_bind_fragment_shader(llvmpipe->draw,
- (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
+ (lp_fs ? lp_fs->draw_data : NULL));
+
+ llvmpipe->fs = lp_fs;
llvmpipe->dirty |= LP_NEW_FS;
}
* Remove shader variant from two lists: the shader's variant list
* and the context's variant list.
*/
-void
+static void
llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
struct lp_fragment_shader_variant *variant)
{
/* Delete draw module's data */
draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
+ if (shader->base.ir.nir)
+ ralloc_free(shader->base.ir.nir);
assert(shader->variants_cached == 0);
FREE((void *) shader->base.tokens);
FREE(shader);
}
if (shader == PIPE_SHADER_VERTEX ||
- shader == PIPE_SHADER_GEOMETRY) {
+ shader == PIPE_SHADER_GEOMETRY ||
+ shader == PIPE_SHADER_TESS_CTRL ||
+ shader == PIPE_SHADER_TESS_EVAL) {
/* Pass the constants to the 'draw' module */
const unsigned size = cb ? cb->buffer_size : 0;
const ubyte *data;
draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
index, data, size);
}
- else {
+ else if (shader == PIPE_SHADER_COMPUTE)
+ llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;
+ else
llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
- }
if (cb && cb->user_buffer) {
pipe_resource_reference(&constants, NULL);
}
}
+static void
+llvmpipe_set_shader_buffers(struct pipe_context *pipe,
+ enum pipe_shader_type shader, unsigned start_slot,
+ unsigned count, const struct pipe_shader_buffer *buffers,
+ unsigned writable_bitmask)
+{
+ struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
+ unsigned i, idx;
+ for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
+ const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;
+
+ util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);
+
+ if (shader == PIPE_SHADER_VERTEX ||
+ shader == PIPE_SHADER_GEOMETRY ||
+ shader == PIPE_SHADER_TESS_CTRL ||
+ shader == PIPE_SHADER_TESS_EVAL) {
+ const unsigned size = buffer ? buffer->buffer_size : 0;
+ const ubyte *data = NULL;
+ if (buffer && buffer->buffer)
+ data = (ubyte *) llvmpipe_resource_data(buffer->buffer);
+ if (data)
+ data += buffer->buffer_offset;
+ draw_set_mapped_shader_buffer(llvmpipe->draw, shader,
+ i, data, size);
+ } else if (shader == PIPE_SHADER_COMPUTE) {
+ llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;
+ } else if (shader == PIPE_SHADER_FRAGMENT) {
+ llvmpipe->dirty |= LP_NEW_FS_SSBOS;
+ }
+ }
+}
+
+static void
+llvmpipe_set_shader_images(struct pipe_context *pipe,
+ enum pipe_shader_type shader, unsigned start_slot,
+ unsigned count, const struct pipe_image_view *images)
+{
+ struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
+ unsigned i, idx;
+
+ draw_flush(llvmpipe->draw);
+ for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
+ const struct pipe_image_view *image = images ? &images[idx] : NULL;
+
+ util_copy_image_view(&llvmpipe->images[shader][i], image);
+ }
+
+ llvmpipe->num_images[shader] = start_slot + count;
+ if (shader == PIPE_SHADER_VERTEX ||
+ shader == PIPE_SHADER_GEOMETRY ||
+ shader == PIPE_SHADER_TESS_CTRL ||
+ shader == PIPE_SHADER_TESS_EVAL) {
+ draw_set_images(llvmpipe->draw,
+ shader,
+ llvmpipe->images[shader],
+ start_slot + count);
+ } else if (shader == PIPE_SHADER_COMPUTE)
+ llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;
+ else
+ llvmpipe->dirty |= LP_NEW_FS_IMAGES;
+}
/**
* Return the blend factor equivalent to a destination alpha of one.
* TODO: there is actually no reason to tie this to context state -- the
* generated code could be cached globally in the screen.
*/
-static void
+static struct lp_fragment_shader_variant_key *
make_variant_key(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
- struct lp_fragment_shader_variant_key *key)
+ char *store)
{
unsigned i;
+ struct lp_fragment_shader_variant_key *key;
+
+ key = (struct lp_fragment_shader_variant_key *)store;
- memset(key, 0, shader->variant_key_size);
+ memset(key, 0, offsetof(struct lp_fragment_shader_variant_key, samplers[1]));
if (lp->framebuffer.zsbuf) {
enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
key->resource_1d = TRUE;
}
+ key->zsbuf_nr_samples = util_res_sample_count(lp->framebuffer.zsbuf->texture);
}
/*
if (lp->rasterizer->clip_halfz) {
key->depth_clamp = 1;
} else {
- key->depth_clamp = (lp->rasterizer->depth_clip == 0) ? 1 : 0;
+ key->depth_clamp = (lp->rasterizer->depth_clip_near == 0) ? 1 : 0;
}
/* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
/* alpha.ref_value is passed in jit_context */
key->flatshade = lp->rasterizer->flatshade;
- if (lp->active_occlusion_queries) {
+ key->multisample = lp->rasterizer->multisample;
+ if (lp->active_occlusion_queries && !lp->queries_disabled) {
key->occlusion_count = TRUE;
}
memcpy(&key->blend, lp->blend, sizeof key->blend);
}
+ key->coverage_samples = 1;
+ 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) {
const struct util_format_description *format_desc;
key->cbuf_format[i] = format;
+ key->cbuf_nr_samples[i] = util_res_sample_count(lp->framebuffer.cbufs[i]->texture);
/*
* Figure out if this is a 1d resource. Note that OpenGL allows crazy
else {
/* no color buffer for this fragment output */
key->cbuf_format[i] = PIPE_FORMAT_NONE;
+ key->cbuf_nr_samples[i] = 0;
blend_rt->colormask = 0x0;
blend_rt->blend_enable = 0;
}
*/
key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
+ struct lp_sampler_static_state *fs_sampler;
+
+ fs_sampler = key->samplers;
+
+ memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);
+
for(i = 0; i < key->nr_samplers; ++i) {
if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
- lp_sampler_static_sampler_state(&key->state[i].sampler_state,
+ lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,
lp->samplers[PIPE_SHADER_FRAGMENT][i]);
}
}
if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
for(i = 0; i < key->nr_sampler_views; ++i) {
- if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
- lp_sampler_static_texture_state(&key->state[i].texture_state,
+ /*
+ * Note sview may exceed what's representable by file_mask.
+ * This will still work, the only downside is that not actually
+ * used views may be included in the shader key.
+ */
+ if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
+ lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
}
}
key->nr_sampler_views = key->nr_samplers;
for(i = 0; i < key->nr_sampler_views; ++i) {
if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
- lp_sampler_static_texture_state(&key->state[i].texture_state,
+ lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
}
}
}
+
+ struct lp_image_static_state *lp_image;
+ lp_image = lp_fs_variant_key_images(key);
+ key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
+ for (i = 0; i < key->nr_images; ++i) {
+ if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
+ lp_sampler_static_texture_state_image(&lp_image[i].image_state,
+ &lp->images[PIPE_SHADER_FRAGMENT][i]);
+ }
+ }
+ return key;
}
llvmpipe_update_fs(struct llvmpipe_context *lp)
{
struct lp_fragment_shader *shader = lp->fs;
- struct lp_fragment_shader_variant_key key;
+ struct lp_fragment_shader_variant_key *key;
struct lp_fragment_shader_variant *variant = NULL;
struct lp_fs_variant_list_item *li;
+ char store[LP_FS_MAX_VARIANT_KEY_SIZE];
- make_variant_key(lp, shader, &key);
+ key = make_variant_key(lp, shader, store);
/* Search the variants for one which matches the key */
li = first_elem(&shader->variants);
while(!at_end(&shader->variants, li)) {
- if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
+ if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
variant = li->base;
break;
}
* Generate the new variant.
*/
t0 = os_time_get();
- variant = generate_variant(lp, shader, &key);
+ variant = generate_variant(lp, shader, key);
t1 = os_time_get();
dt = t1 - t0;
LP_COUNT_ADD(llvm_compile_time, dt);
llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
+
+ llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;
+ llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;
}
-/*
- * 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);
-}
projects / mesa.git / blobdiff