Commit e9fd2a2a authored by Hyunjun Ko's avatar Hyunjun Ko Committed by Marge Bot
Browse files

ir3: Add nonuniform encodings to ir3 encoder and parser



By keeping track of nonuniform access from nir and storing it to ir3.
Signed-off-by: Hyunjun Ko's avatarHyunjun Ko <zzoon@igalia.com>
Reviewed-by: Danylo Piliaiev's avatarDanylo Piliaiev <dpiliaiev@igalia.com>
Reviewed-by: Emma Anholt's avatarEric Anholt <eric@anholt.net>
Part-of: <!9125>
parent 433cdd1c
......@@ -56,6 +56,7 @@ emit_intrinsic_load_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr,
ldib->barrier_class = IR3_BARRIER_BUFFER_R;
ldib->barrier_conflict = IR3_BARRIER_BUFFER_W;
ir3_handle_bindless_cat6(ldib, intr->src[0]);
ir3_handle_nonuniform(ldib, intr);
ir3_split_dest(b, dst, ldib, 0, intr->num_components);
}
......@@ -83,6 +84,7 @@ emit_intrinsic_store_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
stib->barrier_class = IR3_BARRIER_BUFFER_W;
stib->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
ir3_handle_bindless_cat6(stib, intr->src[1]);
ir3_handle_nonuniform(stib, intr);
array_insert(b, b->keeps, stib);
}
......@@ -214,6 +216,7 @@ emit_intrinsic_load_image(struct ir3_context *ctx, nir_intrinsic_instr *intr,
ldib->barrier_class = IR3_BARRIER_IMAGE_R;
ldib->barrier_conflict = IR3_BARRIER_IMAGE_W;
ir3_handle_bindless_cat6(ldib, intr->src[0]);
ir3_handle_nonuniform(ldib, intr);
ir3_split_dest(b, dst, ldib, 0, intr->num_components);
}
......@@ -242,6 +245,7 @@ emit_intrinsic_store_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
stib->barrier_class = IR3_BARRIER_IMAGE_W;
stib->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
ir3_handle_bindless_cat6(stib, intr->src[0]);
ir3_handle_nonuniform(stib, intr);
array_insert(b, b->keeps, stib);
}
......
......@@ -39,6 +39,15 @@
#include "ir3.h"
#include "ir3_context.h"
void
ir3_handle_nonuniform(struct ir3_instruction *instr, nir_intrinsic_instr *intrin)
{
if (nir_intrinsic_has_access(intrin) &&
(nir_intrinsic_access(intrin) & ACCESS_NON_UNIFORM)) {
instr->flags |= IR3_INSTR_NONUNIF;
}
}
void
ir3_handle_bindless_cat6(struct ir3_instruction *instr, nir_src rsrc)
{
......@@ -741,6 +750,7 @@ emit_intrinsic_load_ubo_ldc(struct ir3_context *ctx, nir_intrinsic_instr *intr,
ir3_handle_bindless_cat6(ldc, intr->src[0]);
if (ldc->flags & IR3_INSTR_B)
ctx->so->bindless_ubo = true;
ir3_handle_nonuniform(ldc, intr);
ir3_split_dest(b, dst, ldc, 0, ncomp);
}
......@@ -1233,6 +1243,8 @@ emit_intrinsic_load_image(struct ir3_context *ctx, nir_intrinsic_instr *intr,
sam = emit_sam(ctx, OPC_ISAM, info, type, 0b1111,
ir3_create_collect(ctx, coords, ncoords), NULL);
ir3_handle_nonuniform(sam, intr);
sam->barrier_class = IR3_BARRIER_IMAGE_R;
sam->barrier_conflict = IR3_BARRIER_IMAGE_W;
......@@ -2093,6 +2105,9 @@ get_tex_samp_tex_src(struct ir3_context *ctx, nir_tex_instr *tex)
/* Bindless case */
info.flags |= IR3_INSTR_B;
if (tex->texture_non_uniform || tex->sampler_non_uniform)
info.flags |= IR3_INSTR_NONUNIF;
/* Gather information required to determine which encoding to
* choose as well as for prefetch.
*/
......
......@@ -183,6 +183,7 @@ struct ir3_instruction * ir3_create_collect(struct ir3_context *ctx,
void ir3_split_dest(struct ir3_block *block, struct ir3_instruction **dst,
struct ir3_instruction *src, unsigned base, unsigned n);
void ir3_handle_bindless_cat6(struct ir3_instruction *instr, nir_src rsrc);
void ir3_handle_nonuniform(struct ir3_instruction *instr, nir_intrinsic_instr *intrin);
void emit_intrinsic_image_size_tex(struct ir3_context *ctx, nir_intrinsic_instr *intr,
struct ir3_instruction **dst);
......
......@@ -181,7 +181,12 @@ static void fixup_cat5_s2en(void)
* fix things up.
*/
struct ir3_register *s2en_src = instr->regs[instr->regs_count - 1];
assert(s2en_src->flags & IR3_REG_HALF);
if (instr->flags & IR3_INSTR_B)
assert(!(s2en_src->flags & IR3_REG_HALF));
else
assert(s2en_src->flags & IR3_REG_HALF);
for (int i = 1; i < instr->regs_count - 1; i++) {
instr->regs[i+1] = instr->regs[i];
}
......@@ -919,6 +924,7 @@ cat5_flag: '.' T_3D { instr->flags |= IR3_INSTR_3D; }
| '.' 'p' { instr->flags |= IR3_INSTR_P; }
| '.' 's' { instr->flags |= IR3_INSTR_S; }
| '.' T_S2EN { instr->flags |= IR3_INSTR_S2EN; }
| '.' T_NONUNIFORM { instr->flags |= IR3_INSTR_NONUNIF; }
| '.' T_BASE { instr->flags |= IR3_INSTR_B; instr->cat5.tex_base = $2; }
cat5_flags:
| cat5_flag cat5_flags
......@@ -928,6 +934,7 @@ cat5_tex: T_TEX { if (instr->flags & IR3_INSTR_B) instr->cat5.
cat5_type: '(' type ')' { instr->cat5.type = $2; }
cat5_instr: cat5_opc_dsxypp cat5_flags dst_reg ',' src_reg
| cat5_opc cat5_flags cat5_type dst_reg ',' src_reg ',' src_reg ',' src_reg
| cat5_opc cat5_flags cat5_type dst_reg ',' src_reg ',' src_reg ',' cat5_samp ',' cat5_tex
| cat5_opc cat5_flags cat5_type dst_reg ',' src_reg ',' src_reg ',' cat5_samp
| cat5_opc cat5_flags cat5_type dst_reg ',' src_reg ',' src_reg ',' cat5_tex
......
......@@ -307,6 +307,17 @@ static const struct test {
/* dEQP-GLES31.functional.shaders.opaque_type_indexing.sampler.dynamically_uniform.fragment.sampler2d */
INSTR_6XX(a0c81f07_8100000b, "sam.s2en.uniform (f32)(xyzw)r1.w, r1.y, hr2.x", .parse_fail=true), /* sam.s2en.mode4 (f32)(xyzw)r1.w, r1.y, hr2.x */
/* NonUniform: */
/* dEQP-VK.descriptor_indexing.storage_buffer */
INSTR_6XX(c0260c0a_0a61b180, "ldib.b.untyped.1d.u32.4.nonuniform.base0 r2.z, r2.z, r1.z"),
INSTR_6XX(d0260e0a_09677180, "(sy)stib.b.untyped.1d.u32.4.nonuniform.base0 r2.z, r2.y, r1.w"),
/* dEQP-VK.descriptor_indexing.uniform_texel_buffer */
INSTR_6XX(a0481f00_40000405, "isaml.s2en.nonuniform.base0 (f32)(xyzw)r0.x, r0.z, r0.z, r0.x"),
/* dEQP-VK.descriptor_indexing.storage_image */
INSTR_6XX(d0360c04_02640b81, "(sy)atomic.b.add.typed.2d.u32.1.nonuniform.base0 r1.x, r0.z, r1.z"),
/* dEQP-VK.descriptor_indexing.sampler */
INSTR_6XX(a0c81f00_40000005, "sam.s2en.nonuniform.base0 (f32)(xyzw)r0.x, r0.z, r0.x"),
/* Custom test since we've never seen the blob emit these. */
INSTR_6XX(c0260004_00490000, "getspid.u32 r1.x"),
INSTR_6XX(c0260005_00494000, "getwid.u32 r1.y"),
......
......@@ -178,6 +178,8 @@ extract_cat5_DESC_MODE(struct ir3_instruction *instr)
if (instr->flags & IR3_INSTR_B) {
if (instr->flags & IR3_INSTR_A1EN) {
return CAT5_BINDLESS_A1_UNIFORM;
} else if (instr->flags & IR3_INSTR_NONUNIF) {
return CAT5_BINDLESS_NONUNIFORM;
} else {
return CAT5_BINDLESS_UNIFORM;
}
......
......@@ -58,7 +58,7 @@ SOFTWARE.
The s2en (indirect) or bindless case
</doc>
<display>
{SY}{JP}{NAME}{3D}{A}{O}{P}{S}{S2EN}{UNIFORM}{BASE} {TYPE}({WRMASK}){DST_HALF}{DST}{SRC1}{SRC2}{SRC3}{A1}
{SY}{JP}{NAME}{3D}{A}{O}{P}{S}{S2EN}{UNIFORM}{NONUNIFORM}{BASE} {TYPE}({WRMASK}){DST_HALF}{DST}{SRC1}{SRC2}{SRC3}{A1}
</display>
<field name="BASE_HI" low="19" high="20" type="uint"/>
<field name="SRC3" low="21" high="28" type="#cat5-src3">
......@@ -75,6 +75,7 @@ SOFTWARE.
<derived name="BINDLESS" expr="#cat5-s2enb-is-bindless" type="bool"/>
<derived name="S2EN" expr="#cat5-s2enb-is-indirect" type="bool" display=".s2en"/>
<derived name="UNIFORM" expr="#cat5-s2enb-is-uniform" type="bool" display=".uniform"/>
<derived name="NONUNIFORM" expr="#cat5-s2enb-is-nonuniform" type="bool" display=".nonuniform"/>
<derived name="A1" expr="#cat5-s2enb-uses_a1" type="bool" display=", a1.x"/>
</override>
......@@ -604,6 +605,14 @@ SOFTWARE.
({DESC_MODE} == 5) /* CAT5_BINDLESS_UNIFORM */
</expr>
<!-- Helper to map s2en/bindless DESC_MODE to whether it is non-uniform mode
Note that it returns only for bindless for now, since we need to figure out bindful
uniform/nonuniform mode correctly. See TODO in extract_cat5_DESC_MODE in encode.c -->
<expr name="#cat5-s2enb-is-nonuniform">
({DESC_MODE} == 2) /* CAT5_BINDLESS_NONUNIFORM */ ||
({DESC_MODE} == 3) /* CAT5_BINDLESS_A1_NONUNIFORM */
</expr>
<bitset name="#cat5-src3" size="8">
<doc>bindless/indirect src3, which can either be GPR or samp/tex</doc>
<override expr="#cat5-s2enb-is-indirect">
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment