KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

etnaviv: nir: add compile_check_limits 

To match TGSI compiler behaviour in glmark terrain scene for example.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4199>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4199>
This commit is contained in:
Jonathan Marek 2020-03-15 18:53:36 -04:00 committed by Marge Bot
parent 303842b2db
commit f8bbf44ca4
1 changed files with 73 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

111
src/gallium/drivers/etnaviv/etnaviv_compiler_nir.c
View File

@ -685,6 +685,75 @@ copy_uniform_state_to_shader(struct etna_shader_variant *sobj, uint64_t *consts,
#include "etnaviv_compiler_nir_emit.h"
static bool
etna_compile_check_limits(struct etna_shader_variant *v)
{
const struct etna_specs *specs = v->shader->specs;
int max_uniforms = (v->stage == MESA_SHADER_VERTEX)
? specs->max_vs_uniforms
: specs->max_ps_uniforms;
if (!specs->has_icache && v->needs_icache) {
DBG("Number of instructions (%d) exceeds maximum %d", v->code_size / 4,
specs->max_instructions);
return false;
}
if (v->num_temps > specs->max_registers) {
DBG("Number of registers (%d) exceeds maximum %d", v->num_temps,
specs->max_registers);
return false;
}
if (v->uniforms.imm_count / 4 > max_uniforms) {
DBG("Number of uniforms (%d) exceeds maximum %d",
v->uniforms.imm_count / 4, max_uniforms);
return false;
}
return true;
}
static void
fill_vs_mystery(struct etna_shader_variant *v)
{
const struct etna_specs *specs = v->shader->specs;
v->input_count_unk8 = DIV_ROUND_UP(v->infile.num_reg + 4, 16); /* XXX what is this */
/* fill in "mystery meat" load balancing value. This value determines how
* work is scheduled between VS and PS
* in the unified shader architecture. More precisely, it is determined from
* the number of VS outputs, as well as chip-specific
* vertex output buffer size, vertex cache size, and the number of shader
* cores.
*
* XXX this is a conservative estimate, the "optimal" value is only known for
* sure at link time because some
* outputs may be unused and thus unmapped. Then again, in the general use
* case with GLSL the vertex and fragment
* shaders are linked already before submitting to Gallium, thus all outputs
* are used.
*
* note: TGSI compiler counts all outputs (including position and pointsize), here
* v->outfile.num_reg only counts varyings, +1 to compensate for the position output
* TODO: might have a problem that we don't count pointsize when it is used
*/
int half_out = v->outfile.num_reg / 2 + 1;
assert(half_out);
uint32_t b = ((20480 / (specs->vertex_output_buffer_size -
2 * half_out * specs->vertex_cache_size)) +
9) /
10;
uint32_t a = (b + 256 / (specs->shader_core_count * half_out)) / 2;
v->vs_load_balancing = VIVS_VS_LOAD_BALANCING_A(MIN2(a, 255)) |
VIVS_VS_LOAD_BALANCING_B(MIN2(b, 255)) |
VIVS_VS_LOAD_BALANCING_C(0x3f) |
VIVS_VS_LOAD_BALANCING_D(0x0f);
}
bool
etna_compile_shader_nir(struct etna_shader_variant *v)
{
@ -819,48 +888,14 @@ etna_compile_shader_nir(struct etna_shader_variant *v)
if (s->info.stage == MESA_SHADER_FRAGMENT) {
v->input_count_unk8 = 31; /* XXX what is this */
assert(v->ps_depth_out_reg <= 0);
ralloc_free(c->nir);
FREE(c);
return true;
} else {
fill_vs_mystery(v);
}
v->input_count_unk8 = DIV_ROUND_UP(v->infile.num_reg + 4, 16); /* XXX what is this */
/* fill in "mystery meat" load balancing value. This value determines how
* work is scheduled between VS and PS
* in the unified shader architecture. More precisely, it is determined from
* the number of VS outputs, as well as chip-specific
* vertex output buffer size, vertex cache size, and the number of shader
* cores.
*
* XXX this is a conservative estimate, the "optimal" value is only known for
* sure at link time because some
* outputs may be unused and thus unmapped. Then again, in the general use
* case with GLSL the vertex and fragment
* shaders are linked already before submitting to Gallium, thus all outputs
* are used.
*
* note: TGSI compiler counts all outputs (including position and pointsize), here
* v->outfile.num_reg only counts varyings, +1 to compensate for the position output
* TODO: might have a problem that we don't count pointsize when it is used
*/
int half_out = v->outfile.num_reg / 2 + 1;
assert(half_out);
uint32_t b = ((20480 / (specs->vertex_output_buffer_size -
2 * half_out * specs->vertex_cache_size)) +
9) /
10;
uint32_t a = (b + 256 / (specs->shader_core_count * half_out)) / 2;
v->vs_load_balancing = VIVS_VS_LOAD_BALANCING_A(MIN2(a, 255)) |
VIVS_VS_LOAD_BALANCING_B(MIN2(b, 255)) |
VIVS_VS_LOAD_BALANCING_C(0x3f) |
VIVS_VS_LOAD_BALANCING_D(0x0f);
bool result = etna_compile_check_limits(v);
ralloc_free(c->nir);
FREE(c);
return true;
return result;
}
void