* from the tile buffer after having waited for the scoreboard (which is
* handled by vc4_qpu_emit.c), then do math using your output color and that
* destination value, and update the output color appropriately.
+ *
+ * Once this pass is done, the color write will either have one component (for
+ * single sample) with packed argb8888, or 4 components with the per-sample
+ * argb8888 result.
*/
/**
*/
#include "util/u_format.h"
#include "vc4_qir.h"
-#include "glsl/nir/nir_builder.h"
+#include "compiler/nir/nir_builder.h"
+#include "compiler/nir/nir_format_convert.h"
#include "vc4_context.h"
+static bool
+blend_depends_on_dst_color(struct vc4_compile *c)
+{
+ return (c->fs_key->blend.blend_enable ||
+ c->fs_key->blend.colormask != 0xf ||
+ c->fs_key->logicop_func != PIPE_LOGICOP_COPY);
+}
+
/** Emits a load of the previous fragment color from the tile buffer. */
static nir_ssa_def *
-vc4_nir_get_dst_color(nir_builder *b)
+vc4_nir_get_dst_color(nir_builder *b, int sample)
{
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(b->shader,
nir_intrinsic_load_input);
load->num_components = 1;
- load->const_index[0] = VC4_NIR_TLB_COLOR_READ_INPUT;
- nir_ssa_dest_init(&load->instr, &load->dest, 1, NULL);
+ nir_intrinsic_set_base(load, VC4_NIR_TLB_COLOR_READ_INPUT + sample);
+ load->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
+ nir_ssa_dest_init(&load->instr, &load->dest, 1, 32, NULL);
nir_builder_instr_insert(b, &load->instr);
return &load->dest.ssa;
}
-static nir_ssa_def *
-vc4_nir_srgb_decode(nir_builder *b, nir_ssa_def *srgb)
-{
- nir_ssa_def *is_low = nir_flt(b, srgb, nir_imm_float(b, 0.04045));
- nir_ssa_def *low = nir_fmul(b, srgb, nir_imm_float(b, 1.0 / 12.92));
- nir_ssa_def *high = nir_fpow(b,
- nir_fmul(b,
- nir_fadd(b, srgb,
- nir_imm_float(b, 0.055)),
- nir_imm_float(b, 1.0 / 1.055)),
- nir_imm_float(b, 2.4));
-
- return nir_bcsel(b, is_low, low, high);
-}
-
-static nir_ssa_def *
-vc4_nir_srgb_encode(nir_builder *b, nir_ssa_def *linear)
-{
- nir_ssa_def *is_low = nir_flt(b, linear, nir_imm_float(b, 0.0031308));
- nir_ssa_def *low = nir_fmul(b, linear, nir_imm_float(b, 12.92));
- nir_ssa_def *high = nir_fsub(b,
- nir_fmul(b,
- nir_imm_float(b, 1.055),
- nir_fpow(b,
- linear,
- nir_imm_float(b, 0.41666))),
- nir_imm_float(b, 0.055));
-
- return nir_bcsel(b, is_low, low, high);
-}
-
static nir_ssa_def *
vc4_blend_channel_f(nir_builder *b,
nir_ssa_def **src,
return nir_imm_float(b, 1.0);
}
case PIPE_BLENDFACTOR_CONST_COLOR:
- return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_X + channel);
+ return nir_load_system_value(b,
+ nir_intrinsic_load_blend_const_color_r_float +
+ channel,
+ 0, 32);
case PIPE_BLENDFACTOR_CONST_ALPHA:
- return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_W);
+ return nir_load_blend_const_color_a_float(b);
case PIPE_BLENDFACTOR_ZERO:
return nir_imm_float(b, 0.0);
case PIPE_BLENDFACTOR_INV_SRC_COLOR:
return nir_fsub(b, nir_imm_float(b, 1.0), dst[channel]);
case PIPE_BLENDFACTOR_INV_CONST_COLOR:
return nir_fsub(b, nir_imm_float(b, 1.0),
- vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_X + channel));
+ nir_load_system_value(b,
+ nir_intrinsic_load_blend_const_color_r_float +
+ channel,
+ 0, 32));
case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
return nir_fsub(b, nir_imm_float(b, 1.0),
- vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_W));
+ nir_load_blend_const_color_a_float(b));
default:
case PIPE_BLENDFACTOR_SRC1_COLOR:
nir_imm_int(b, ~0),
a_chan);
case PIPE_BLENDFACTOR_CONST_COLOR:
- return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_RGBA);
+ return nir_load_blend_const_color_rgba8888_unorm(b);
case PIPE_BLENDFACTOR_CONST_ALPHA:
- return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_AAAA);
+ return nir_load_blend_const_color_aaaa8888_unorm(b);
case PIPE_BLENDFACTOR_ZERO:
return nir_imm_int(b, 0);
case PIPE_BLENDFACTOR_INV_SRC_COLOR:
case PIPE_BLENDFACTOR_INV_DST_COLOR:
return nir_inot(b, dst);
case PIPE_BLENDFACTOR_INV_CONST_COLOR:
- return nir_inot(b, vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_RGBA));
+ return nir_inot(b,
+ nir_load_blend_const_color_rgba8888_unorm(b));
case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
- return nir_inot(b, vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_AAAA));
+ return nir_inot(b,
+ nir_load_blend_const_color_aaaa8888_unorm(b));
default:
case PIPE_BLENDFACTOR_SRC1_COLOR:
}
}
-static nir_ssa_def *
-vc4_nir_pipe_compare_func(nir_builder *b, int func,
- nir_ssa_def *src0, nir_ssa_def *src1)
-{
- switch (func) {
- default:
- fprintf(stderr, "Unknown compare func %d\n", func);
- /* FALLTHROUGH */
- case PIPE_FUNC_NEVER:
- return nir_imm_int(b, 0);
- case PIPE_FUNC_ALWAYS:
- return nir_imm_int(b, ~0);
- case PIPE_FUNC_EQUAL:
- return nir_feq(b, src0, src1);
- case PIPE_FUNC_NOTEQUAL:
- return nir_fne(b, src0, src1);
- case PIPE_FUNC_GREATER:
- return nir_flt(b, src1, src0);
- case PIPE_FUNC_GEQUAL:
- return nir_fge(b, src0, src1);
- case PIPE_FUNC_LESS:
- return nir_flt(b, src0, src1);
- case PIPE_FUNC_LEQUAL:
- return nir_fge(b, src1, src0);
- }
-}
-
-static void
-vc4_nir_emit_alpha_test_discard(struct vc4_compile *c, nir_builder *b,
- nir_ssa_def *alpha)
-{
- if (!c->fs_key->alpha_test)
- return;
-
- nir_ssa_def *alpha_ref =
- vc4_nir_get_state_uniform(b, QUNIFORM_ALPHA_REF);
- nir_ssa_def *condition =
- vc4_nir_pipe_compare_func(b, c->fs_key->alpha_test_func,
- alpha, alpha_ref);
-
- nir_intrinsic_instr *discard =
- nir_intrinsic_instr_create(b->shader,
- nir_intrinsic_discard_if);
- discard->num_components = 1;
- discard->src[0] = nir_src_for_ssa(nir_inot(b, condition));
- nir_builder_instr_insert(b, &discard->instr);
-}
-
static nir_ssa_def *
vc4_nir_swizzle_and_pack(struct vc4_compile *c, nir_builder *b,
nir_ssa_def **colors)
}
-static void
-vc4_nir_lower_blend_instr(struct vc4_compile *c, nir_builder *b,
- nir_intrinsic_instr *intr)
+static nir_ssa_def *
+vc4_nir_blend_pipeline(struct vc4_compile *c, nir_builder *b, nir_ssa_def *src,
+ int sample)
{
enum pipe_format color_format = c->fs_key->color_format;
const uint8_t *format_swiz = vc4_get_format_swizzle(color_format);
bool srgb = util_format_is_srgb(color_format);
/* Pull out the float src/dst color components. */
- nir_ssa_def *packed_dst_color = vc4_nir_get_dst_color(b);
+ nir_ssa_def *packed_dst_color = vc4_nir_get_dst_color(b, sample);
nir_ssa_def *dst_vec4 = nir_unpack_unorm_4x8(b, packed_dst_color);
nir_ssa_def *src_color[4], *unpacked_dst_color[4];
for (unsigned i = 0; i < 4; i++) {
- src_color[i] = nir_channel(b, intr->src[0].ssa, i);
+ src_color[i] = nir_channel(b, src, i);
unpacked_dst_color[i] = nir_channel(b, dst_vec4, i);
}
- vc4_nir_emit_alpha_test_discard(c, b, src_color[3]);
+ if (c->fs_key->sample_alpha_to_one && c->fs_key->msaa)
+ src_color[3] = nir_imm_float(b, 1.0);
nir_ssa_def *packed_color;
if (srgb) {
/* Turn dst color to linear. */
for (int i = 0; i < 3; i++)
- dst_color[i] = vc4_nir_srgb_decode(b, dst_color[i]);
+ dst_color[i] = nir_format_srgb_to_linear(b, dst_color[i]);
nir_ssa_def *blend_color[4];
vc4_do_blending_f(c, b, blend_color, src_color, dst_color);
/* sRGB encode the output color */
for (int i = 0; i < 3; i++)
- blend_color[i] = vc4_nir_srgb_encode(b, blend_color[i]);
+ blend_color[i] = nir_format_linear_to_srgb(b, blend_color[i]);
packed_color = vc4_nir_swizzle_and_pack(c, b, blend_color);
} else {
colormask &= ~(0xff << (i * 8));
}
}
- packed_color = nir_ior(b,
- nir_iand(b, packed_color,
- nir_imm_int(b, colormask)),
- nir_iand(b, packed_dst_color,
- nir_imm_int(b, ~colormask)));
- /* Turn the old vec4 output into a store of the packed color. */
- nir_instr_rewrite_src(&intr->instr, &intr->src[0],
- nir_src_for_ssa(packed_color));
+ return nir_ior(b,
+ nir_iand(b, packed_color,
+ nir_imm_int(b, colormask)),
+ nir_iand(b, packed_dst_color,
+ nir_imm_int(b, ~colormask)));
+}
+
+static int
+vc4_nir_next_output_driver_location(nir_shader *s)
+{
+ int maxloc = -1;
+
+ nir_foreach_variable(var, &s->outputs)
+ maxloc = MAX2(maxloc, (int)var->data.driver_location);
+
+ return maxloc + 1;
+}
+
+static void
+vc4_nir_store_sample_mask(struct vc4_compile *c, nir_builder *b,
+ nir_ssa_def *val)
+{
+ nir_variable *sample_mask = nir_variable_create(c->s, nir_var_shader_out,
+ glsl_uint_type(),
+ "sample_mask");
+ sample_mask->data.driver_location =
+ vc4_nir_next_output_driver_location(c->s);
+ sample_mask->data.location = FRAG_RESULT_SAMPLE_MASK;
+
+ nir_intrinsic_instr *intr =
+ nir_intrinsic_instr_create(c->s, nir_intrinsic_store_output);
intr->num_components = 1;
+ nir_intrinsic_set_base(intr, sample_mask->data.driver_location);
+
+ intr->src[0] = nir_src_for_ssa(val);
+ intr->src[1] = nir_src_for_ssa(nir_imm_int(b, 0));
+ nir_builder_instr_insert(b, &intr->instr);
}
-static bool
-vc4_nir_lower_blend_block(nir_block *block, void *state)
+static void
+vc4_nir_lower_blend_instr(struct vc4_compile *c, nir_builder *b,
+ nir_intrinsic_instr *intr)
{
- struct vc4_compile *c = state;
+ nir_ssa_def *frag_color = intr->src[0].ssa;
+
+ if (c->fs_key->sample_alpha_to_coverage) {
+ nir_ssa_def *a = nir_channel(b, frag_color, 3);
+
+ /* XXX: We should do a nice dither based on the fragment
+ * coordinate, instead.
+ */
+ nir_ssa_def *num_samples = nir_imm_float(b, VC4_MAX_SAMPLES);
+ nir_ssa_def *num_bits = nir_f2i32(b, nir_fmul(b, a, num_samples));
+ nir_ssa_def *bitmask = nir_isub(b,
+ nir_ishl(b,
+ nir_imm_int(b, 1),
+ num_bits),
+ nir_imm_int(b, 1));
+ vc4_nir_store_sample_mask(c, b, bitmask);
+ }
- nir_foreach_instr(block, instr) {
+ /* The TLB color read returns each sample in turn, so if our blending
+ * depends on the destination color, we're going to have to run the
+ * blending function separately for each destination sample value, and
+ * then output the per-sample color using TLB_COLOR_MS.
+ */
+ nir_ssa_def *blend_output;
+ if (c->fs_key->msaa && blend_depends_on_dst_color(c)) {
+ c->msaa_per_sample_output = true;
+
+ nir_ssa_def *samples[4];
+ for (int i = 0; i < VC4_MAX_SAMPLES; i++)
+ samples[i] = vc4_nir_blend_pipeline(c, b, frag_color, i);
+ blend_output = nir_vec4(b,
+ samples[0], samples[1],
+ samples[2], samples[3]);
+ } else {
+ blend_output = vc4_nir_blend_pipeline(c, b, frag_color, 0);
+ }
+
+ nir_instr_rewrite_src(&intr->instr, &intr->src[0],
+ nir_src_for_ssa(blend_output));
+ intr->num_components = blend_output->num_components;
+}
+
+static bool
+vc4_nir_lower_blend_block(nir_block *block, struct vc4_compile *c)
+{
+ nir_foreach_instr_safe(instr, block) {
if (instr->type != nir_instr_type_intrinsic)
continue;
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
nir_variable *output_var = NULL;
nir_foreach_variable(var, &c->s->outputs) {
- if (var->data.driver_location == intr->const_index[0]) {
+ if (var->data.driver_location ==
+ nir_intrinsic_base(intr)) {
output_var = var;
break;
}
}
void
-vc4_nir_lower_blend(struct vc4_compile *c)
+vc4_nir_lower_blend(nir_shader *s, struct vc4_compile *c)
{
- nir_foreach_overload(c->s, overload) {
- if (overload->impl) {
- nir_foreach_block(overload->impl,
- vc4_nir_lower_blend_block, c);
+ nir_foreach_function(function, s) {
+ if (function->impl) {
+ nir_foreach_block(block, function->impl) {
+ vc4_nir_lower_blend_block(block, c);
+ }
- nir_metadata_preserve(overload->impl,
+ nir_metadata_preserve(function->impl,
nir_metadata_block_index |
nir_metadata_dominance);
}
}
+
+ /* If we didn't do alpha-to-coverage on the output color, we still
+ * need to pass glSampleMask() through.
+ */
+ if (c->fs_key->sample_coverage && !c->fs_key->sample_alpha_to_coverage) {
+ nir_function_impl *impl = nir_shader_get_entrypoint(s);
+ nir_builder b;
+ nir_builder_init(&b, impl);
+ b.cursor = nir_after_block(nir_impl_last_block(impl));
+
+ vc4_nir_store_sample_mask(c, &b, nir_load_sample_mask_in(&b));
+ }
}