From: Marek Olšák <[email protected]>
---
src/amd/common/ac_llvm_build.c | 19 +++++++++++++++++--
src/amd/common/ac_llvm_build.h | 1 +
src/amd/common/ac_nir_to_llvm.c | 2 +-
.../drivers/radeonsi/si_shader_tgsi_mem.c | 4 +---
4 files changed, 20 insertions(+), 6 deletions(-)
diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index fc6dc396d38..ed510a34d6f 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -551,20 +551,36 @@ LLVMValueRef ac_build_expand_to_vec4(struct
ac_llvm_context *ctx,
}
elemtype = LLVMTypeOf(value);
}
while (num_channels < 4)
chan[num_channels++] = LLVMGetUndef(elemtype);
return ac_build_gather_values(ctx, chan, 4);
}
+LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
+{
+ unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
+ const char *name;
+
+ if (type_size == 2)
+ name = "llvm.rint.f16";
+ else if (type_size == 4)
+ name = "llvm.rint.f32";
+ else
+ name = "llvm.rint.f64";
+
+ return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
+ AC_FUNC_ATTR_READNONE);
+}
+
LLVMValueRef
ac_build_fdiv(struct ac_llvm_context *ctx,
LLVMValueRef num,
LLVMValueRef den)
{
/* If we do (num / den), LLVM >= 7.0 does:
* return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f
: 1.0f));
*
* If we do (num * (1 / den)), LLVM does:
* return num * v_rcp_f32(den);
@@ -729,22 +745,21 @@ ac_prepare_cube_coords(struct ac_llvm_context *ctx,
LLVMValueRef *coords_arg,
LLVMValueRef *derivs_arg)
{
LLVMBuilderRef builder = ctx->builder;
struct cube_selection_coords selcoords;
LLVMValueRef coords[3];
LLVMValueRef invma;
if (is_array && !is_lod) {
- LLVMValueRef tmp = coords_arg[3];
- tmp = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &tmp,
1, 0);
+ LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
/* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
*
* "For Array forms, the array layer used will be
*
* max(0, min(d−1, floor(layer+0.5)))
*
* where d is the depth of the texture array and layer
* comes from the component indicated in the tables below.
* Workaroudn for an issue where the layer is taken from a
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 83aad02183e..32d62450dfe 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -168,20 +168,21 @@ ac_build_gather_values_extended(struct ac_llvm_context
*ctx,
unsigned value_stride,
bool load,
bool always_vector);
LLVMValueRef
ac_build_gather_values(struct ac_llvm_context *ctx,
LLVMValueRef *values,
unsigned value_count);
LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
LLVMValueRef value,
unsigned num_channels);
+LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
LLVMValueRef
ac_build_fdiv(struct ac_llvm_context *ctx,
LLVMValueRef num,
LLVMValueRef den);
LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
LLVMValueRef num,
LLVMValueRef multiplier,
LLVMValueRef pre_shift,
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 312383db36c..ffc64a79d95 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -3304,21 +3304,21 @@ static void tex_fetch_ptrs(struct ac_nir_context *ctx,
}
if (fmask_ptr && (instr->op == nir_texop_txf_ms ||
instr->op == nir_texop_samples_identical))
*fmask_ptr = get_sampler_desc(ctx, texture_deref_instr,
AC_DESC_FMASK, instr, false, false);
}
static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
LLVMValueRef coord)
{
coord = ac_to_float(ctx, coord);
- coord = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &coord, 1,
0);
+ coord = ac_build_round(ctx, coord);
coord = ac_to_integer(ctx, coord);
return coord;
}
static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
{
LLVMValueRef result = NULL;
struct ac_image_args args = { 0 };
LLVMValueRef fmask_ptr = NULL, sample_index = NULL;
LLVMValueRef ddx = NULL, ddy = NULL;
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index cabc448a082..8c44831bccb 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -1439,23 +1439,21 @@ static void build_tex_intrinsic(const struct
lp_build_tgsi_action *action,
opcode == TGSI_OPCODE_TXD,
target == TGSI_TEXTURE_CUBE_ARRAY ||
target == TGSI_TEXTURE_SHADOWCUBE_ARRAY,
opcode == TGSI_OPCODE_LODQ,
args.coords, args.derivs);
} else if (tgsi_is_array_sampler(target) &&
opcode != TGSI_OPCODE_TXF &&
opcode != TGSI_OPCODE_TXF_LZ &&
ctx->screen->info.chip_class <= VI) {
unsigned array_coord = target == TGSI_TEXTURE_1D_ARRAY ? 1 : 2;
- args.coords[array_coord] =
- ac_build_intrinsic(&ctx->ac, "llvm.rint.f32", ctx->f32,
- &args.coords[array_coord], 1, 0);
+ args.coords[array_coord] = ac_build_round(&ctx->ac,
args.coords[array_coord]);
}
/* 1D textures are allocated and used as 2D on GFX9. */
if (ctx->screen->info.chip_class >= GFX9) {
LLVMValueRef filler;
/* Use 0.5, so that we don't sample the border color. */
if (opcode == TGSI_OPCODE_TXF ||
opcode == TGSI_OPCODE_TXF_LZ)
filler = ctx->i32_0;
--
2.17.1
_______________________________________________
mesa-dev mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
