mesa: Move ATI_fragment_shader fog code emit to a NIR lowering pass.

Now it's implemented as a RMW of the FRAG_RESULT_COLOR output var (or
adjusting the store_output intrinsic's value for lowered i/o), which
should be reusable other places we might want to emit shader code for fog
(ARB_fp, fixed function fragment shaders).

Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23111>
This commit is contained in:
Emma Anholt 2023-05-16 12:38:57 -07:00 committed by Marge Bot
parent b26a9efc5a
commit 66951679f2
5 changed files with 159 additions and 63 deletions

View file

@ -358,6 +358,7 @@ files_libmesa = files(
'state_tracker/st_nir.h',
'state_tracker/st_nir_builtins.c',
'state_tracker/st_nir_lower_builtin.c',
'state_tracker/st_nir_lower_fog.c',
'state_tracker/st_nir_lower_tex_src_plane.c',
'state_tracker/st_pbo.c',
'state_tracker/st_pbo_compute.c',

View file

@ -32,9 +32,6 @@
#include "st_atifs_to_nir.h"
#include "compiler/nir/nir_builder.h"
#define FOG_PARAMS_UNIFORM (MAX_NUM_FRAGMENT_CONSTANTS_ATI + 0)
#define FOG_COLOR_UNIFORM (MAX_NUM_FRAGMENT_CONSTANTS_ATI + 1)
/**
* Intermediate state used during shader translation.
*/
@ -119,8 +116,7 @@ load_input(struct st_translate *t, gl_varying_slot slot)
{
if (!t->inputs[slot]) {
nir_variable *var = nir_create_variable_with_location(t->b->shader, nir_var_shader_in, slot,
slot == VARYING_SLOT_FOGC ?
glsl_float_type() : glsl_vec4_type());
glsl_vec4_type());
var->data.interpolation = INTERP_MODE_NONE;
t->inputs[slot] = nir_load_var(t->b, var);
@ -418,8 +414,7 @@ compile_instruction(struct st_translate *t,
}
/* Creates the uniform variable referencing the ATI_fragment_shader constants
* plus the optimized fog state.
/* Creates the uniform variable referencing the ATI_fragment_shader constants.
*/
static void
st_atifs_setup_uniforms(struct st_translate *t, struct gl_program *program)
@ -475,51 +470,8 @@ st_translate_atifs_program(struct ati_fragment_shader *atifs,
}
}
if (t->regs_written[atifs->NumPasses-1][0]) {
nir_ssa_def *color = t->temps[0];
if (key->fog) {
nir_ssa_def *fogc = load_input(t, VARYING_SLOT_FOGC);
nir_ssa_def *params = atifs_load_uniform(t, FOG_PARAMS_UNIFORM);
/* compute the 1 component fog factor f */
nir_ssa_def *f = NULL;
if (key->fog == FOG_LINEAR) {
f = nir_ffma(t->b, fogc,
nir_channel(t->b, params, 0),
nir_channel(t->b, params, 1));
} else if (key->fog == FOG_EXP) {
/* EXP formula: f = exp(-dens * z)
* with optimized parameters:
* f = MUL(fogcoord, oparams.z); f= EX2(-f)
*/
f = nir_fmul(t->b, fogc, nir_channel(t->b, params, 2));
f = nir_fexp2(t->b, nir_fneg(t->b, f));
} else if (key->fog == FOG_EXP2) {
/* EXP2 formula: f = exp(-(dens * z)^2)
* with optimized parameters:
* f = MUL(fogcoord, oparams.w); f=MUL(f, f); f= EX2(-f)
*/
f = nir_fmul(t->b, fogc, nir_channel(t->b, params, 3));
f = nir_fmul(t->b, f, f);
f = nir_fexp2(t->b, nir_fneg(t->b, f));
}
f = nir_fsat(t->b, f);
nir_ssa_def *fog_color = nir_flrp(t->b,
atifs_load_uniform(t, FOG_COLOR_UNIFORM),
color,
f);
color = nir_vec4(t->b,
nir_channel(t->b, fog_color, 0),
nir_channel(t->b, fog_color, 1),
nir_channel(t->b, fog_color, 2),
nir_channel(t->b, color, 3));
}
nir_store_var(t->b, t->fragcolor, color, 0xf);
}
if (t->regs_written[atifs->NumPasses-1][0])
nir_store_var(t->b, t->fragcolor, t->temps[0], 0xf);
return b.shader;
}
@ -536,11 +488,6 @@ st_init_atifs_prog(struct gl_context *ctx, struct gl_program *prog)
unsigned pass, i, r, optype, arg;
static const gl_state_index16 fog_params_state[STATE_LENGTH] =
{STATE_FOG_PARAMS_OPTIMIZED, 0, 0};
static const gl_state_index16 fog_color[STATE_LENGTH] =
{STATE_FOG_COLOR, 0, 0, 0};
prog->info.inputs_read = 0;
prog->info.outputs_written = BITFIELD64_BIT(FRAG_RESULT_COLOR);
prog->SamplersUsed = 0;
@ -587,17 +534,13 @@ st_init_atifs_prog(struct gl_context *ctx, struct gl_program *prog)
}
}
}
/* we may need fog */
prog->info.inputs_read |= BITFIELD64_BIT(VARYING_SLOT_FOGC);
/* we always have the ATI_fs constants, and the fog params */
/* we always have the ATI_fs constants */
for (i = 0; i < MAX_NUM_FRAGMENT_CONSTANTS_ATI; i++) {
_mesa_add_parameter(prog->Parameters, PROGRAM_UNIFORM,
NULL, 4, GL_FLOAT, NULL, NULL, true);
}
ASSERTED uint32_t ref;
ref = _mesa_add_state_reference(prog->Parameters, fog_params_state);
assert(ref == FOG_PARAMS_UNIFORM);
ref = _mesa_add_state_reference(prog->Parameters, fog_color);
assert(ref == FOG_COLOR_UNIFORM);
}

View file

@ -87,6 +87,9 @@ st_nir_state_variable_create(struct nir_shader *shader,
const struct glsl_type *type,
const gl_state_index16 state[STATE_LENGTH]);
bool st_nir_lower_fog(struct nir_shader *s, enum gl_fog_mode fog_mode,
struct gl_program_parameter_list *paramList);
#ifdef __cplusplus
}
#endif

View file

@ -0,0 +1,137 @@
/*
* Copyright © 2023 Google LLC
* SPDX-License-Identifier: MIT
*/
#include "nir_builder.h"
#include "nir_builtin_builder.h"
#include "st_nir.h"
static nir_ssa_def *
fog_result(nir_builder *b, nir_ssa_def *color, enum gl_fog_mode fog_mode, struct gl_program_parameter_list *paramList)
{
nir_shader *s = b->shader;
nir_variable *fogc_var =
nir_create_variable_with_location(s, nir_var_shader_in, VARYING_SLOT_FOGC, glsl_float_type());
nir_ssa_def *fogc = nir_load_var(b, fogc_var);
s->info.inputs_read |= VARYING_BIT_FOGC;
static const gl_state_index16 fog_params_tokens[STATE_LENGTH] = {STATE_FOG_PARAMS_OPTIMIZED};
static const gl_state_index16 fog_color_tokens[STATE_LENGTH] = {STATE_FOG_COLOR};
nir_variable *fog_params_var = st_nir_state_variable_create(s, glsl_vec4_type(), fog_params_tokens);
fog_params_var->data.driver_location = _mesa_add_state_reference(paramList, fog_params_tokens);
nir_ssa_def *params = nir_load_var(b, fog_params_var);
nir_variable *fog_color_var = st_nir_state_variable_create(s, glsl_vec4_type(), fog_color_tokens);
fog_color_var->data.driver_location = _mesa_add_state_reference(paramList, fog_color_tokens);
nir_ssa_def *fog_color = nir_load_var(b, fog_color_var);
/* compute the 1 component fog factor f */
nir_ssa_def *f = NULL;
switch (fog_mode) {
case FOG_LINEAR:
/* f = (end - z) / (end - start)
*
* gl_MesaFogParamsOptimized gives us (-1 / (end - start)) and
* (end / (end - start)) so we can generate a single MAD.
*/
f = nir_fmad(b, fogc,
nir_channel(b, params, 0),
nir_channel(b, params, 1));
break;
case FOG_EXP:
/* f = e^(-(density * fogcoord))
*
* gl_MesaFogParamsOptimized gives us density/ln(2) so we can
* use EXP2 which is generally the native instruction without
* having to do any further math on the fog density uniform.
*/
f = nir_fmul(b, fogc, nir_channel(b, params, 2));
f = nir_fexp2(b, nir_fneg(b, f));
break;
case FOG_EXP2:
/* f = e^(-(density * fogcoord)^2)
*
* gl_MesaFogParamsOptimized gives us density/sqrt(ln(2)) so we
* can do this like FOG_EXP but with a squaring after the
* multiply by density.
*/
f = nir_fmul(b, fogc, nir_channel(b, params, 3));
f = nir_fmul(b, f, f);
f = nir_fexp2(b, nir_fneg(b, f));
break;
default:
unreachable("unsupported fog mode");
}
f = nir_fsat(b, f);
/* Not using flrp because we may end up lowering fog after driver lowering
* that meant to remove all lrps.
*/
return nir_fmad(b, color, f, nir_fmul(b, fog_color, nir_fsub_imm(b, 1.0, f)));
}
struct lower_fog_state {
enum gl_fog_mode fog_mode;
struct gl_program_parameter_list *paramList;
};
static bool
st_nir_lower_fog_instr(nir_builder *b, nir_instr *instr, void *_state)
{
const struct lower_fog_state *state = _state;
if (instr->type != nir_instr_type_intrinsic)
return false;
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
if (intr->intrinsic != nir_intrinsic_store_output)
return false;
int loc = nir_intrinsic_io_semantics(intr).location;
if (loc != FRAG_RESULT_COLOR)
return false;
b->cursor = nir_before_instr(instr);
nir_ssa_def *color = nir_ssa_for_src(b, intr->src[0], intr->num_components);
color = nir_resize_vector(b, color, 4);
nir_ssa_def *fog = fog_result(b, color, state->fog_mode, state->paramList);
/* retain the non-fog-blended alpha value for color */
color = nir_vector_insert_imm(b, fog, nir_channel(b, color, 3), 3);
nir_instr_rewrite_src_ssa(instr, &intr->src[0], nir_resize_vector(b, color, intr->num_components));
return true;
}
bool
st_nir_lower_fog(nir_shader *s, enum gl_fog_mode fog_mode, struct gl_program_parameter_list *paramList)
{
if (s->info.io_lowered) {
struct lower_fog_state state = {
.fog_mode = fog_mode,
.paramList = paramList,
};
nir_shader_instructions_pass(s, st_nir_lower_fog_instr,
nir_metadata_block_index |
nir_metadata_dominance,
&state);
} else {
nir_variable *color_var = nir_find_variable_with_location(s, nir_var_shader_out, FRAG_RESULT_COLOR);
nir_builder b;
nir_builder_init(&b, nir_shader_get_entrypoint(s));
b.cursor = nir_after_block(nir_impl_last_block(b.impl));
nir_ssa_def *color = nir_load_var(&b, color_var);
color = fog_result(&b, color, fog_mode, paramList);
nir_store_var(&b, color_var, color, 0x7);
nir_metadata_preserve(b.impl, nir_metadata_block_index |
nir_metadata_dominance); }
return true;
}

View file

@ -53,6 +53,7 @@
#include "tgsi/tgsi_dump.h"
#include "tgsi/tgsi_parse.h"
#include "tgsi/tgsi_ureg.h"
#include "nir_builder.h"
#include "nir/nir_to_tgsi.h"
#include "util/u_memory.h"
@ -926,6 +927,17 @@ st_create_fp_variant(struct st_context *st,
bool finalize = false;
if (fp->ati_fs) {
if (key->fog) {
NIR_PASS_V(state.ir.nir, st_nir_lower_fog, key->fog, fp->Parameters);
NIR_PASS_V(state.ir.nir, nir_lower_io_to_temporaries,
nir_shader_get_entrypoint(state.ir.nir),
true, false);
nir_lower_global_vars_to_local(state.ir.nir);
}
finalize = true;
}
if (key->clamp_color) {
NIR_PASS_V(state.ir.nir, nir_lower_clamp_color_outputs);
finalize = true;