[Mesa-dev] [PATCH] gallivm: implement better control of per-quad/per-element/scalar lod
sroland at vmware.com
sroland at vmware.com
Thu Aug 15 18:07:47 PDT 2013
From: Roland Scheidegger <sroland at vmware.com>
There's a new debug value used to disable per-quad lod optimizations
in fragment shader (ignored for vs/gs as the results are just too wrong
typically). Also trying to detect if a supplied lod value is really a
scalar (if it's coming from immediate or constant file) in which case
sampler code can use this to stay on per-quad-lod path (in fact for
explicit lod could simplify even further and use same lod for both
quads in the avx case but this is not implemented yet).
Still need to actually implement per-element lod bias (and derivatives),
and need to handle per-element lod in size queries.
---
src/gallium/auxiliary/draw/draw_llvm_sample.c | 8 +-
src/gallium/auxiliary/gallivm/lp_bld_debug.h | 3 +-
src/gallium/auxiliary/gallivm/lp_bld_init.c | 1 +
src/gallium/auxiliary/gallivm/lp_bld_sample.h | 11 +-
src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c | 8 +-
src/gallium/auxiliary/gallivm/lp_bld_tgsi.h | 5 +-
src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c | 158 ++++++++++++++++-----
src/gallium/drivers/llvmpipe/lp_tex_sample.c | 8 +-
8 files changed, 147 insertions(+), 55 deletions(-)
diff --git a/src/gallium/auxiliary/draw/draw_llvm_sample.c b/src/gallium/auxiliary/draw/draw_llvm_sample.c
index 97b0255..a6341fa 100644
--- a/src/gallium/auxiliary/draw/draw_llvm_sample.c
+++ b/src/gallium/auxiliary/draw/draw_llvm_sample.c
@@ -238,7 +238,7 @@ draw_llvm_sampler_soa_emit_fetch_texel(const struct lp_build_sampler_soa *base,
const struct lp_derivatives *derivs,
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef *texel)
{
struct draw_llvm_sampler_soa *sampler = (struct draw_llvm_sampler_soa *)base;
@@ -257,7 +257,7 @@ draw_llvm_sampler_soa_emit_fetch_texel(const struct lp_build_sampler_soa *base,
coords,
offsets,
derivs,
- lod_bias, explicit_lod, scalar_lod,
+ lod_bias, explicit_lod, lod_property,
texel);
}
@@ -272,7 +272,7 @@ draw_llvm_sampler_soa_emit_size_query(const struct lp_build_sampler_soa *base,
unsigned texture_unit,
unsigned target,
boolean is_sviewinfo,
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef explicit_lod, /* optional */
LLVMValueRef *sizes_out)
{
@@ -287,7 +287,7 @@ draw_llvm_sampler_soa_emit_size_query(const struct lp_build_sampler_soa *base,
texture_unit,
target,
is_sviewinfo,
- scalar_lod,
+ lod_property,
explicit_lod,
sizes_out);
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_debug.h b/src/gallium/auxiliary/gallivm/lp_bld_debug.h
index 4f38edf..76c39af 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_debug.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_debug.h
@@ -43,7 +43,8 @@
#define GALLIVM_DEBUG_PERF (1 << 4)
#define GALLIVM_DEBUG_NO_BRILINEAR (1 << 5)
#define GALLIVM_DEBUG_NO_RHO_APPROX (1 << 6)
-#define GALLIVM_DEBUG_GC (1 << 7)
+#define GALLIVM_DEBUG_NO_QUAD_LOD (1 << 7)
+#define GALLIVM_DEBUG_GC (1 << 8)
#ifdef __cplusplus
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.c b/src/gallium/auxiliary/gallivm/lp_bld_init.c
index e4cc058..61eadb8 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_init.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c
@@ -80,6 +80,7 @@ static const struct debug_named_value lp_bld_debug_flags[] = {
{ "perf", GALLIVM_DEBUG_PERF, NULL },
{ "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
{ "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
+ { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
{ "gc", GALLIVM_DEBUG_GC, NULL },
DEBUG_NAMED_VALUE_END
};
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.h b/src/gallium/auxiliary/gallivm/lp_bld_sample.h
index 6d8fe88..a6901d1 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.h
@@ -61,6 +61,13 @@ struct lp_derivatives
};
+enum lp_sampler_lod_property {
+ lp_sampler_lod_scalar,
+ lp_sampler_lod_perelement,
+ lp_sampler_lod_perquad
+};
+
+
/**
* Texture static state.
*
@@ -476,7 +483,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm,
const struct lp_derivatives *derivs,
LLVMValueRef lod_bias,
LLVMValueRef explicit_lod,
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef texel_out[4]);
@@ -497,7 +504,7 @@ lp_build_size_query_soa(struct gallivm_state *gallivm,
unsigned texture_unit,
unsigned target,
boolean is_sviewinfo,
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef explicit_lod,
LLVMValueRef *sizes_out);
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index 20a08cb..76de006 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -1646,7 +1646,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm,
const struct lp_derivatives *derivs, /* optional */
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef texel_out[4])
{
unsigned target = static_texture_state->target;
@@ -1733,7 +1733,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm,
* There are other situations where at least the multiple int lods could be
* avoided like min and max lod being equal.
*/
- if (explicit_lod && !scalar_lod &&
+ if (explicit_lod && lod_property == lp_sampler_lod_perelement &&
((is_fetch && target != PIPE_BUFFER) ||
(!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
bld.num_lods = type.length;
@@ -1925,7 +1925,7 @@ lp_build_sample_soa(struct gallivm_state *gallivm,
bld4.levelf_type.length = 1;
bld4.leveli_type = lp_int_type(bld4.levelf_type);
- if (explicit_lod && !scalar_lod &&
+ if (explicit_lod && lod_property == lp_sampler_lod_perelement &&
((is_fetch && target != PIPE_BUFFER) ||
(!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
bld4.num_lods = type4.length;
@@ -2046,7 +2046,7 @@ lp_build_size_query_soa(struct gallivm_state *gallivm,
unsigned texture_unit,
unsigned target,
boolean is_sviewinfo,
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef explicit_lod,
LLVMValueRef *sizes_out)
{
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
index 9d27f5f..522302e 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
@@ -40,6 +40,7 @@
#include "gallivm/lp_bld.h"
#include "gallivm/lp_bld_tgsi_action.h"
#include "gallivm/lp_bld_limits.h"
+#include "gallivm/lp_bld_sample.h"
#include "lp_bld_type.h"
#include "pipe/p_compiler.h"
#include "pipe/p_state.h"
@@ -184,7 +185,7 @@ struct lp_build_sampler_soa
const struct lp_derivatives *derivs,
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
- boolean scalar_lod,
+ enum lp_sampler_lod_property,
LLVMValueRef *texel);
void
@@ -194,7 +195,7 @@ struct lp_build_sampler_soa
unsigned unit,
unsigned target,
boolean need_nr_mips,
- boolean scalar_lod,
+ enum lp_sampler_lod_property,
LLVMValueRef explicit_lod, /* optional */
LLVMValueRef *sizes_out);
};
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
index cab53df..a873394 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
@@ -1601,6 +1601,54 @@ tgsi_to_pipe_tex_target(unsigned tgsi_target)
}
}
+
+static enum lp_sampler_lod_property
+lp_build_lod_property(
+ struct lp_build_tgsi_context *bld_base,
+ const struct tgsi_full_instruction *inst,
+ unsigned src_op,
+ LLVMValueRef fetched_src)
+{
+ const struct tgsi_full_src_register *reg = &inst->Src[src_op];
+ enum lp_sampler_lod_property lod_property;
+
+ /*
+ * Not much we can do here. We could try catching inputs declared
+ * with constant interpolation and similar but it seems pointless.
+ * What seems to be of more value would be to recognize temps holding
+ * broadcasted scalars but no way we can do it.
+ * Tried asking llvm but without any success (using LLVMIsConstant
+ * even though this isn't exactly what we'd need), even as simple as
+ * IMM[0] UINT32 (0,-1,0,0)
+ * MOV TEMP[0] IMM[0].yyyy
+ * SVIEWINFO TEMP[1], TEMP[0].xxxx, SVIEWINFO[0]
+ * doesn't work.
+ * This means there's ZERO chance this will ever catch a scalar lod
+ * with traditional tex opcodes as well as texel fetches, since the lod
+ * comes from the same reg as coords (except some test shaders).
+ * There's at least hope for sample opcodes as well as size queries.
+ */
+ if (reg->Register.File == TGSI_FILE_CONSTANT ||
+ reg->Register.File == TGSI_FILE_IMMEDIATE/* ||
+ LLVMIsConstant(fetched_src)*/) {
+ lod_property = lp_sampler_lod_scalar;
+ }
+ else if (bld_base->info->processor == TGSI_PROCESSOR_FRAGMENT) {
+ if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
+ lod_property = lp_sampler_lod_perelement;
+ }
+ else {
+ lod_property = lp_sampler_lod_perquad;
+ }
+ }
+ else {
+ /* never use scalar (per-quad) lod the results are just too wrong. */
+ lod_property = lp_sampler_lod_perelement;
+ }
+ return lod_property;
+}
+
+
/**
* High-level instruction translators.
*/
@@ -1618,7 +1666,7 @@ emit_tex( struct lp_build_tgsi_soa_context *bld,
LLVMValueRef offsets[3] = { NULL };
struct lp_derivatives derivs;
struct lp_derivatives *deriv_ptr = NULL;
- boolean scalar_lod;
+ enum lp_sampler_lod_property lod_property = lp_sampler_lod_scalar;
unsigned num_derivs, num_offsets, i;
unsigned shadow_coord = 0;
unsigned layer_coord = 0;
@@ -1690,13 +1738,18 @@ emit_tex( struct lp_build_tgsi_soa_context *bld,
}
/* Note lod and especially projected are illegal in a LOT of cases */
- if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
- lod_bias = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
- explicit_lod = NULL;
- }
- else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
- lod_bias = NULL;
- explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
+ modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ LLVMValueRef lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
+ lod_bias = lod;
+ explicit_lod = NULL;
+ }
+ else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ lod_bias = NULL;
+ explicit_lod = lod;
+ }
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0, lod);
}
else {
lod_bias = NULL;
@@ -1738,6 +1791,21 @@ emit_tex( struct lp_build_tgsi_soa_context *bld,
}
deriv_ptr = &derivs;
unit = inst->Src[3].Register.Index;
+ /*
+ * could also check all src regs if constant but I doubt such
+ * cases exist in practice.
+ */
+ if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
+ if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
+ lod_property = lp_sampler_lod_perelement;
+ }
+ else {
+ lod_property = lp_sampler_lod_perquad;
+ }
+ }
+ else {
+ lod_property = lp_sampler_lod_perelement;
+ }
} else {
unit = inst->Src[1].Register.Index;
}
@@ -1750,9 +1818,6 @@ emit_tex( struct lp_build_tgsi_soa_context *bld,
}
}
- /* TODO: use scalar lod if explicit_lod, lod_bias or derivs are broadcasted scalars */
- scalar_lod = bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT;
-
bld->sampler->emit_fetch_texel(bld->sampler,
bld->bld_base.base.gallivm,
bld->bld_base.base.type,
@@ -1761,7 +1826,7 @@ emit_tex( struct lp_build_tgsi_soa_context *bld,
coords,
offsets,
deriv_ptr,
- lod_bias, explicit_lod, scalar_lod,
+ lod_bias, explicit_lod, lod_property,
texel);
}
@@ -1779,7 +1844,8 @@ emit_sample(struct lp_build_tgsi_soa_context *bld,
LLVMValueRef offsets[3] = { NULL };
struct lp_derivatives derivs;
struct lp_derivatives *deriv_ptr = NULL;
- boolean scalar_lod;
+ enum lp_sampler_lod_property lod_property = lp_sampler_lod_scalar;
+
unsigned num_offsets, num_derivs, i;
unsigned layer_coord = 0;
@@ -1841,13 +1907,18 @@ emit_sample(struct lp_build_tgsi_soa_context *bld,
return;
}
- if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
- lod_bias = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
- explicit_lod = NULL;
- }
- else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
- lod_bias = NULL;
- explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
+ modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ LLVMValueRef lod = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
+ lod_bias = lod;
+ explicit_lod = NULL;
+ }
+ else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ lod_bias = NULL;
+ explicit_lod = lod;
+ }
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0, lod);
}
else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
lod_bias = NULL;
@@ -1885,6 +1956,21 @@ emit_sample(struct lp_build_tgsi_soa_context *bld,
derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 4, dim);
}
deriv_ptr = &derivs;
+ /*
+ * could also check all src regs if constant but I doubt such
+ * cases exist in practice.
+ */
+ if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
+ if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
+ lod_property = lp_sampler_lod_perelement;
+ }
+ else {
+ lod_property = lp_sampler_lod_perquad;
+ }
+ }
+ else {
+ lod_property = lp_sampler_lod_perelement;
+ }
}
/* some advanced gather instructions (txgo) would require 4 offsets */
@@ -1895,10 +1981,6 @@ emit_sample(struct lp_build_tgsi_soa_context *bld,
}
}
- /* TODO: use scalar lod if explicit_lod, lod_bias or derivs are broadcasted scalars */
- scalar_lod = bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT ||
- modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO;
-
bld->sampler->emit_fetch_texel(bld->sampler,
bld->bld_base.base.gallivm,
bld->bld_base.base.type,
@@ -1907,7 +1989,7 @@ emit_sample(struct lp_build_tgsi_soa_context *bld,
coords,
offsets,
deriv_ptr,
- lod_bias, explicit_lod, scalar_lod,
+ lod_bias, explicit_lod, lod_property,
texel);
if (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
@@ -1935,7 +2017,7 @@ emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
LLVMValueRef explicit_lod = NULL;
LLVMValueRef coords[3];
LLVMValueRef offsets[3] = { NULL };
- boolean scalar_lod;
+ enum lp_sampler_lod_property lod_property = lp_sampler_lod_scalar;
unsigned dims, i;
unsigned layer_coord = 0;
@@ -1984,6 +2066,7 @@ emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
/* always have lod except for buffers ? */
if (target != TGSI_TEXTURE_BUFFER) {
explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0, explicit_lod);
}
for (i = 0; i < dims; i++) {
@@ -2002,9 +2085,6 @@ emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
}
}
- /* TODO: use scalar lod if explicit_lod is broadcasted scalar */
- scalar_lod = bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT;
-
bld->sampler->emit_fetch_texel(bld->sampler,
bld->bld_base.base.gallivm,
bld->bld_base.base.type,
@@ -2013,7 +2093,7 @@ emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
coords,
offsets,
NULL,
- NULL, explicit_lod, scalar_lod,
+ NULL, explicit_lod, lod_property,
texel);
if (is_samplei &&
@@ -2038,7 +2118,7 @@ emit_size_query( struct lp_build_tgsi_soa_context *bld,
boolean is_sviewinfo)
{
LLVMValueRef explicit_lod;
- boolean scalar_lod;
+ enum lp_sampler_lod_property lod_property;
unsigned has_lod;
unsigned i;
unsigned unit = inst->Src[1].Register.Index;
@@ -2068,22 +2148,24 @@ emit_size_query( struct lp_build_tgsi_soa_context *bld,
return;
}
- if (has_lod)
- explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 0 );
- else
+ if (has_lod) {
+ explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 0);
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0, explicit_lod);
+ }
+ else {
explicit_lod = NULL;
+ lod_property = lp_sampler_lod_scalar;
+ }
- pipe_target = tgsi_to_pipe_tex_target(target);
- /* TODO: use scalar lod if explicit_lod is broadcasted scalar */
- scalar_lod = bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT;
+ pipe_target = tgsi_to_pipe_tex_target(target);
bld->sampler->emit_size_query(bld->sampler,
bld->bld_base.base.gallivm,
bld->bld_base.int_bld.type,
unit, pipe_target,
is_sviewinfo,
- scalar_lod,
+ lod_property,
explicit_lod,
sizes_out);
}
diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample.c b/src/gallium/drivers/llvmpipe/lp_tex_sample.c
index 68a4be9..f0a4a34 100644
--- a/src/gallium/drivers/llvmpipe/lp_tex_sample.c
+++ b/src/gallium/drivers/llvmpipe/lp_tex_sample.c
@@ -244,7 +244,7 @@ lp_llvm_sampler_soa_emit_fetch_texel(const struct lp_build_sampler_soa *base,
const struct lp_derivatives *derivs,
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef *texel)
{
struct lp_llvm_sampler_soa *sampler = (struct lp_llvm_sampler_soa *)base;
@@ -268,7 +268,7 @@ lp_llvm_sampler_soa_emit_fetch_texel(const struct lp_build_sampler_soa *base,
coords,
offsets,
derivs,
- lod_bias, explicit_lod, scalar_lod,
+ lod_bias, explicit_lod, lod_property,
texel);
}
@@ -282,7 +282,7 @@ lp_llvm_sampler_soa_emit_size_query(const struct lp_build_sampler_soa *base,
unsigned texture_unit,
unsigned target,
boolean is_sviewinfo,
- boolean scalar_lod,
+ enum lp_sampler_lod_property lod_property,
LLVMValueRef explicit_lod, /* optional */
LLVMValueRef *sizes_out)
{
@@ -297,7 +297,7 @@ lp_llvm_sampler_soa_emit_size_query(const struct lp_build_sampler_soa *base,
texture_unit,
target,
is_sviewinfo,
- scalar_lod,
+ lod_property,
explicit_lod,
sizes_out);
}
--
1.7.9.5
More information about the mesa-dev
mailing list