OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ac: replace glc,slc with cache_policy for loads

Browse Source 
cosmetic change

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
This commit is contained in:
Marek Olšák
2019-06-28 20:53:15 -04:00
parent a29e781961
commit 8a71f60194
7 changed files with 79 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
src/amd/common/ac_llvm_build.c
View File

@@ -1107,13 +1107,11 @@ LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
return ac_build_load_custom(ctx, base_ptr, index, true, true, false); return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
} }
static LLVMValueRef get_cache_policy(struct ac_llvm_context *ctx, static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
bool load, bool glc, bool slc) unsigned cache_policy)
{ {
return LLVMConstInt(ctx->i32, return cache_policy |
(glc ? ac_glc : 0) + (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
(slc ? ac_slc : 0) +
(ctx->chip_class >= GFX10 && glc && load ? ac_dlc : 0), 0);
} }
static void static void
@@ -1302,8 +1300,7 @@ ac_build_llvm7_buffer_load_common(struct ac_llvm_context *ctx,
LLVMValueRef vindex, LLVMValueRef vindex,
LLVMValueRef voffset, LLVMValueRef voffset,
unsigned num_channels, unsigned num_channels,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate, bool can_speculate,
bool use_format) bool use_format)
{ {
@@ -1311,8 +1308,8 @@ ac_build_llvm7_buffer_load_common(struct ac_llvm_context *ctx,
LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""), LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
vindex ? vindex : ctx->i32_0, vindex ? vindex : ctx->i32_0,
voffset, voffset,
LLVMConstInt(ctx->i1, glc, 0), LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
LLVMConstInt(ctx->i1, slc, 0) LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), 0)
}; };
unsigned func = CLAMP(num_channels, 1, 3) - 1; unsigned func = CLAMP(num_channels, 1, 3) - 1;
@@ -1341,8 +1338,7 @@ ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
LLVMValueRef soffset, LLVMValueRef soffset,
unsigned num_channels, unsigned num_channels,
LLVMTypeRef channel_type, LLVMTypeRef channel_type,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate, bool can_speculate,
bool use_format, bool use_format,
bool structurized) bool structurized)
@@ -1354,7 +1350,7 @@ ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
args[idx++] = vindex ? vindex : ctx->i32_0; args[idx++] = vindex ? vindex : ctx->i32_0;
args[idx++] = voffset ? voffset : ctx->i32_0; args[idx++] = voffset ? voffset : ctx->i32_0;
args[idx++] = soffset ? soffset : ctx->i32_0; args[idx++] = soffset ? soffset : ctx->i32_0;
args[idx++] = get_cache_policy(ctx, true, glc, slc); args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels; unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
const char *indexing_kind = structurized ? "struct" : "raw"; const char *indexing_kind = structurized ? "struct" : "raw";
char name[256], type_name[8]; char name[256], type_name[8];
@@ -1382,8 +1378,7 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
unsigned inst_offset, unsigned inst_offset,
unsigned glc, unsigned cache_policy,
unsigned slc,
bool can_speculate, bool can_speculate,
bool allow_smem) bool allow_smem)
{ {
@@ -1393,8 +1388,8 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
if (soffset) if (soffset)
offset = LLVMBuildAdd(ctx->builder, offset, soffset, ""); offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
if (allow_smem && !slc && if (allow_smem && !(cache_policy & ac_slc) &&
(!glc || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= GFX8))) { (!(cache_policy & ac_glc) || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= GFX8))) {
assert(vindex == NULL); assert(vindex == NULL);
LLVMValueRef result[8]; LLVMValueRef result[8];
@@ -1411,7 +1406,7 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
LLVMValueRef args[3] = { LLVMValueRef args[3] = {
rsrc, rsrc,
offset, offset,
get_cache_policy(ctx, true, glc, false), LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
}; };
result[i] = ac_build_intrinsic(ctx, intrname, result[i] = ac_build_intrinsic(ctx, intrname,
ctx->f32, args, num_args, ctx->f32, args, num_args,
@@ -1430,13 +1425,13 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
offset, ctx->i32_0, offset, ctx->i32_0,
num_channels, ctx->f32, num_channels, ctx->f32,
glc, slc, cache_policy,
can_speculate, false, can_speculate, false,
false); false);
} }
return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, offset, return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, offset,
num_channels, glc, slc, num_channels, cache_policy,
can_speculate, false); can_speculate, false);
} }
@@ -1445,17 +1440,16 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
LLVMValueRef vindex, LLVMValueRef vindex,
LLVMValueRef voffset, LLVMValueRef voffset,
unsigned num_channels, unsigned num_channels,
bool glc, unsigned cache_policy,
bool can_speculate) bool can_speculate)
{ {
if (HAVE_LLVM >= 0x800) { if (HAVE_LLVM >= 0x800) {
return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0, return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
num_channels, ctx->f32, num_channels, ctx->f32,
glc, false, cache_policy, can_speculate, true, true);
can_speculate, true, true);
} }
return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, voffset, return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, voffset,
num_channels, glc, false, num_channels, cache_policy,
can_speculate, true); can_speculate, true);
} }
@@ -1464,14 +1458,13 @@ LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
LLVMValueRef vindex, LLVMValueRef vindex,
LLVMValueRef voffset, LLVMValueRef voffset,
unsigned num_channels, unsigned num_channels,
bool glc, unsigned cache_policy,
bool can_speculate) bool can_speculate)
{ {
if (HAVE_LLVM >= 0x800) { if (HAVE_LLVM >= 0x800) {
return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0, return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
num_channels, ctx->f32, num_channels, ctx->f32,
glc, false, cache_policy, can_speculate, true, true);
can_speculate, true, true);
} }
LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), ""); LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
@@ -1486,7 +1479,7 @@ LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
LLVMConstInt(ctx->i32, 2, 0), ""); LLVMConstInt(ctx->i32, 2, 0), "");
return ac_build_llvm7_buffer_load_common(ctx, new_rsrc, vindex, voffset, return ac_build_llvm7_buffer_load_common(ctx, new_rsrc, vindex, voffset,
num_channels, glc, false, num_channels, cache_policy,
can_speculate, true); can_speculate, true);
} }
@@ -1542,8 +1535,7 @@ ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
unsigned num_channels, unsigned num_channels,
unsigned dfmt, unsigned dfmt,
unsigned nfmt, unsigned nfmt,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate, bool can_speculate,
bool structurized) bool structurized)
{ {
@@ -1555,7 +1547,7 @@ ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
args[idx++] = voffset ? voffset : ctx->i32_0; args[idx++] = voffset ? voffset : ctx->i32_0;
args[idx++] = soffset ? soffset : ctx->i32_0; args[idx++] = soffset ? soffset : ctx->i32_0;
args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0); args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0);
args[idx++] = get_cache_policy(ctx, true, glc, slc); args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels; unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
const char *indexing_kind = structurized ? "struct" : "raw"; const char *indexing_kind = structurized ? "struct" : "raw";
char name[256], type_name[8]; char name[256], type_name[8];
@@ -1580,8 +1572,7 @@ ac_build_tbuffer_load(struct ac_llvm_context *ctx,
unsigned num_channels, unsigned num_channels,
unsigned dfmt, unsigned dfmt,
unsigned nfmt, unsigned nfmt,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate, bool can_speculate,
bool structurized) /* only matters for LLVM 8+ */ bool structurized) /* only matters for LLVM 8+ */
{ {
@@ -1590,7 +1581,7 @@ ac_build_tbuffer_load(struct ac_llvm_context *ctx,
return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset, return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
soffset, num_channels, soffset, num_channels,
dfmt, nfmt, glc, slc, dfmt, nfmt, cache_policy,
can_speculate, structurized); can_speculate, structurized);
} }
@@ -1602,8 +1593,8 @@ ac_build_tbuffer_load(struct ac_llvm_context *ctx,
immoffset, immoffset,
LLVMConstInt(ctx->i32, dfmt, false), LLVMConstInt(ctx->i32, dfmt, false),
LLVMConstInt(ctx->i32, nfmt, false), LLVMConstInt(ctx->i32, nfmt, false),
LLVMConstInt(ctx->i1, glc, false), LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), false),
LLVMConstInt(ctx->i1, slc, false), LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), false),
}; };
unsigned func = CLAMP(num_channels, 1, 3) - 1; unsigned func = CLAMP(num_channels, 1, 3) - 1;
LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32}; LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
@@ -1627,13 +1618,12 @@ ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
unsigned num_channels, unsigned num_channels,
unsigned dfmt, unsigned dfmt,
unsigned nfmt, unsigned nfmt,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate) bool can_speculate)
{ {
return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset, return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
immoffset, num_channels, dfmt, nfmt, glc, immoffset, num_channels, dfmt, nfmt,
slc, can_speculate, true); cache_policy, can_speculate, true);
} }
LLVMValueRef LLVMValueRef
@@ -1645,13 +1635,12 @@ ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
unsigned num_channels, unsigned num_channels,
unsigned dfmt, unsigned dfmt,
unsigned nfmt, unsigned nfmt,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate) bool can_speculate)
{ {
return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset, return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
immoffset, num_channels, dfmt, nfmt, glc, immoffset, num_channels, dfmt, nfmt,
slc, can_speculate, false); cache_policy, can_speculate, false);
} }
LLVMValueRef LLVMValueRef
@@ -1660,7 +1649,7 @@ ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
LLVMValueRef immoffset, LLVMValueRef immoffset,
bool glc) unsigned cache_policy)
{ {
LLVMValueRef res; LLVMValueRef res;
@@ -1670,14 +1659,14 @@ ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */ /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL, res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
voffset, soffset, voffset, soffset,
1, ctx->i16, glc, false, 1, ctx->i16, cache_policy,
false, false, false); false, false, false);
} else { } else {
unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16; unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT; unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset, res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
immoffset, 1, dfmt, nfmt, glc, false, immoffset, 1, dfmt, nfmt, cache_policy,
false); false);
res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, ""); res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
@@ -1692,7 +1681,7 @@ ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
LLVMValueRef immoffset, LLVMValueRef immoffset,
bool glc) unsigned cache_policy)
{ {
LLVMValueRef res; LLVMValueRef res;
@@ -1702,14 +1691,14 @@ ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */ /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL, res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
voffset, soffset, voffset, soffset,
1, ctx->i8, glc, false, 1, ctx->i8, cache_policy,
false, false, false); false, false, false);
} else { } else {
unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8; unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT; unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset, res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
immoffset, 1, dfmt, nfmt, glc, false, immoffset, 1, dfmt, nfmt, cache_policy,
false); false);
res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, ""); res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
@@ -1811,8 +1800,7 @@ ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
LLVMValueRef vindex, LLVMValueRef vindex,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate) bool can_speculate)
{ {
LLVMValueRef tmp; LLVMValueRef tmp;
@@ -1851,13 +1839,13 @@ ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2); unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
loads[i] = ac_build_llvm8_buffer_load_common( loads[i] = ac_build_llvm8_buffer_load_common(
ctx, rsrc, vindex, voffset, tmp, ctx, rsrc, vindex, voffset, tmp,
num_channels, channel_type, glc, slc, num_channels, channel_type, cache_policy,
can_speculate, false, true); can_speculate, false, true);
} else { } else {
tmp = LLVMBuildAdd(ctx->builder, voffset, tmp, ""); tmp = LLVMBuildAdd(ctx->builder, voffset, tmp, "");
loads[i] = ac_build_llvm7_buffer_load_common( loads[i] = ac_build_llvm7_buffer_load_common(
ctx, rsrc, vindex, tmp, ctx, rsrc, vindex, tmp,
1 << (load_log_size - 2), glc, slc, can_speculate, false); 1 << (load_log_size - 2), cache_policy, can_speculate, false);
} }
if (load_log_size >= 2) if (load_log_size >= 2)
loads[i] = ac_to_integer(ctx, loads[i]); loads[i] = ac_to_integer(ctx, loads[i]);

20
src/amd/common/ac_llvm_build.h
View File

@@ -288,8 +288,7 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
unsigned inst_offset, unsigned inst_offset,
unsigned glc, unsigned cache_policy,
unsigned slc,
bool can_speculate, bool can_speculate,
bool allow_smem); bool allow_smem);
@@ -298,7 +297,7 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
LLVMValueRef vindex, LLVMValueRef vindex,
LLVMValueRef voffset, LLVMValueRef voffset,
unsigned num_channels, unsigned num_channels,
bool glc, unsigned cache_policy,
bool can_speculate); bool can_speculate);
/* load_format that handles the stride & element count better if idxen is /* load_format that handles the stride & element count better if idxen is
@@ -308,7 +307,7 @@ LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
LLVMValueRef vindex, LLVMValueRef vindex,
LLVMValueRef voffset, LLVMValueRef voffset,
unsigned num_channels, unsigned num_channels,
bool glc, unsigned cache_policy,
bool can_speculate); bool can_speculate);
LLVMValueRef LLVMValueRef
@@ -317,7 +316,7 @@ ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
LLVMValueRef immoffset, LLVMValueRef immoffset,
bool glc); unsigned cache_policy);
LLVMValueRef LLVMValueRef
ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx, ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
@@ -325,7 +324,7 @@ ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
LLVMValueRef immoffset, LLVMValueRef immoffset,
bool glc); unsigned cache_policy);
LLVMValueRef LLVMValueRef
ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx, ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
@@ -337,8 +336,7 @@ ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
unsigned num_channels, unsigned num_channels,
unsigned dfmt, unsigned dfmt,
unsigned nfmt, unsigned nfmt,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate); bool can_speculate);
LLVMValueRef LLVMValueRef
@@ -350,8 +348,7 @@ ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
unsigned num_channels, unsigned num_channels,
unsigned dfmt, unsigned dfmt,
unsigned nfmt, unsigned nfmt,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate); bool can_speculate);
/* For ac_build_fetch_format. /* For ac_build_fetch_format.
@@ -380,8 +377,7 @@ ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
LLVMValueRef vindex, LLVMValueRef vindex,
LLVMValueRef voffset, LLVMValueRef voffset,
LLVMValueRef soffset, LLVMValueRef soffset,
bool glc, unsigned cache_policy,
bool slc,
bool can_speculate); bool can_speculate);
void void

21
src/amd/common/ac_nir_to_llvm.c
View File

@@ -1303,14 +1303,14 @@ static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
args->coords[0], args->coords[0],
ctx->ac.i32_0, ctx->ac.i32_0,
util_last_bit(mask), util_last_bit(mask),
false, true); 0, true);
} else { } else {
return ac_build_buffer_load_format(&ctx->ac, return ac_build_buffer_load_format(&ctx->ac,
args->resource, args->resource,
args->coords[0], args->coords[0],
ctx->ac.i32_0, ctx->ac.i32_0,
util_last_bit(mask), util_last_bit(mask),
false, true); 0, true);
} }
} }
@@ -1743,22 +1743,21 @@ static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
offset, offset,
ctx->ac.i32_0, ctx->ac.i32_0,
immoffset, immoffset,
cache_policy & ac_glc); cache_policy);
} else if (load_bytes == 2) { } else if (load_bytes == 2) {
ret = ac_build_tbuffer_load_short(&ctx->ac, ret = ac_build_tbuffer_load_short(&ctx->ac,
rsrc, rsrc,
offset, offset,
ctx->ac.i32_0, ctx->ac.i32_0,
immoffset, immoffset,
cache_policy & ac_glc); cache_policy);
} else { } else {
int num_channels = util_next_power_of_two(load_bytes) / 4; int num_channels = util_next_power_of_two(load_bytes) / 4;
bool can_speculate = access & ACCESS_CAN_REORDER; bool can_speculate = access & ACCESS_CAN_REORDER;
ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels, ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
vindex, offset, immoffset, 0, vindex, offset, immoffset, 0,
cache_policy & ac_glc, 0, cache_policy, can_speculate, false);
can_speculate, false);
} }
LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret))); LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
@@ -1804,7 +1803,7 @@ static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
offset, offset,
ctx->ac.i32_0, ctx->ac.i32_0,
immoffset, immoffset,
false); 0);
} else { } else {
assert(load_bytes == 2); assert(load_bytes == 2);
results[i] = ac_build_tbuffer_load_short(&ctx->ac, results[i] = ac_build_tbuffer_load_short(&ctx->ac,
@@ -1812,13 +1811,13 @@ static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
offset, offset,
ctx->ac.i32_0, ctx->ac.i32_0,
immoffset, immoffset,
false); 0);
} }
} }
ret = ac_build_gather_values(&ctx->ac, results, num_components); ret = ac_build_gather_values(&ctx->ac, results, num_components);
} else { } else {
ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset, ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
NULL, 0, false, false, true, true); NULL, 0, 0, true, true);
ret = ac_trim_vector(&ctx->ac, ret, num_components); ret = ac_trim_vector(&ctx->ac, ret, num_components);
} }
@@ -2467,7 +2466,7 @@ static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
bool can_speculate = access & ACCESS_CAN_REORDER; bool can_speculate = access & ACCESS_CAN_REORDER;
res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex, res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
ctx->ac.i32_0, num_channels, ctx->ac.i32_0, num_channels,
!!(args.cache_policy & ac_glc), args.cache_policy,
can_speculate); can_speculate);
res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels); res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
@@ -3476,7 +3475,7 @@ static LLVMValueRef get_bindless_index_from_uniform(struct ac_nir_context *ctx,
LLVMValueRef ubo_index = ctx->abi->load_ubo(ctx->abi, ctx->ac.i32_0); LLVMValueRef ubo_index = ctx->abi->load_ubo(ctx->abi, ctx->ac.i32_0);
LLVMValueRef ret = ac_build_buffer_load(&ctx->ac, ubo_index, 1, NULL, offset, LLVMValueRef ret = ac_build_buffer_load(&ctx->ac, ubo_index, 1, NULL, offset,
NULL, 0, false, false, true, true); NULL, 0, 0, true, true);
return LLVMBuildBitCast(ctx->ac.builder, ret, ctx->ac.i32, ""); return LLVMBuildBitCast(ctx->ac.builder, ret, ctx->ac.i32, "");
} }

9
src/amd/vulkan/radv_nir_to_llvm.c
View File

@@ -1630,7 +1630,7 @@ load_tes_input(struct ac_shader_abi *abi,
buf_addr = LLVMBuildAdd(ctx->ac.builder, buf_addr, comp_offset, ""); buf_addr = LLVMBuildAdd(ctx->ac.builder, buf_addr, comp_offset, "");
result = ac_build_buffer_load(&ctx->ac, ctx->hs_ring_tess_offchip, num_components, NULL, result = ac_build_buffer_load(&ctx->ac, ctx->hs_ring_tess_offchip, num_components, NULL,
buf_addr, ctx->oc_lds, is_compact ? (4 * const_index) : 0, 1, 0, true, false); buf_addr, ctx->oc_lds, is_compact ? (4 * const_index) : 0, ac_glc, true, false);
result = ac_trim_vector(&ctx->ac, result, num_components); result = ac_trim_vector(&ctx->ac, result, num_components);
return result; return result;
} }
@@ -1673,7 +1673,7 @@ load_gs_input(struct ac_shader_abi *abi,
ctx->esgs_ring, 1, ctx->esgs_ring, 1,
ctx->ac.i32_0, ctx->ac.i32_0,
vtx_offset, soffset, vtx_offset, soffset,
0, 1, 0, true, false); 0, ac_glc, true, false);
} }
if (ac_get_type_size(type) == 2) { if (ac_get_type_size(type) == 2) {
@@ -2236,8 +2236,7 @@ handle_vs_input_decl(struct radv_shader_context *ctx,
LLVMConstInt(ctx->ac.i32, attrib_offset, false), LLVMConstInt(ctx->ac.i32, attrib_offset, false),
ctx->ac.i32_0, ctx->ac.i32_0, ctx->ac.i32_0, ctx->ac.i32_0,
num_channels, num_channels,
data_format, num_format, data_format, num_format, 0, true);
false, false, true);
if (ctx->options->key.vs.post_shuffle & (1 << attrib_index)) { if (ctx->options->key.vs.post_shuffle & (1 << attrib_index)) {
LLVMValueRef c[4]; LLVMValueRef c[4];
@@ -4041,7 +4040,7 @@ ac_gs_copy_shader_emit(struct radv_shader_context *ctx)
ctx->gsvs_ring[0], ctx->gsvs_ring[0],
1, ctx->ac.i32_0, 1, ctx->ac.i32_0,
vtx_offset, soffset, vtx_offset, soffset,
0, 1, 1, true, false); 0, ac_glc | ac_slc, true, false);
LLVMTypeRef type = LLVMGetAllocatedType(ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]); LLVMTypeRef type = LLVMGetAllocatedType(ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
if (ac_get_type_size(type) == 2) { if (ac_get_type_size(type) == 2) {

2
src/gallium/drivers/radeonsi/si_compute_prim_discard.c
View File

@@ -473,7 +473,7 @@ void si_build_prim_discard_compute_shader(struct si_shader_context *ctx)
for (unsigned i = 0; i < 3; i++) { for (unsigned i = 0; i < 3; i++) {
index[i] = ac_build_buffer_load_format(&ctx->ac, input_indexbuf, index[i] = ac_build_buffer_load_format(&ctx->ac, input_indexbuf,
index[i], ctx->i32_0, 1, index[i], ctx->i32_0, 1,
false, true); 0, true);
index[i] = ac_to_integer(&ctx->ac, index[i]); index[i] = ac_to_integer(&ctx->ac, index[i]);
} }
} }

21
src/gallium/drivers/radeonsi/si_shader.c
View File

@@ -541,8 +541,7 @@ void si_llvm_load_input_vs(
tmp = ac_build_opencoded_load_format( tmp = ac_build_opencoded_load_format(
&ctx->ac, fix_fetch.u.log_size, fix_fetch.u.num_channels_m1 + 1, &ctx->ac, fix_fetch.u.log_size, fix_fetch.u.num_channels_m1 + 1,
fix_fetch.u.format, fix_fetch.u.reverse, !opencode, fix_fetch.u.format, fix_fetch.u.reverse, !opencode,
t_list, vertex_index, ctx->ac.i32_0, ctx->ac.i32_0, t_list, vertex_index, ctx->ac.i32_0, ctx->ac.i32_0, 0, true);
false, false, true);
for (unsigned i = 0; i < 4; ++i) for (unsigned i = 0; i < 4; ++i)
out[i] = LLVMBuildExtractElement(ctx->ac.builder, tmp, LLVMConstInt(ctx->i32, i, false), ""); out[i] = LLVMBuildExtractElement(ctx->ac.builder, tmp, LLVMConstInt(ctx->i32, i, false), "");
return; return;
@@ -568,7 +567,7 @@ void si_llvm_load_input_vs(
for (unsigned i = 0; i < num_fetches; ++i) { for (unsigned i = 0; i < num_fetches; ++i) {
LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0); LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
fetches[i] = ac_build_buffer_load_format(&ctx->ac, t_list, vertex_index, voffset, fetches[i] = ac_build_buffer_load_format(&ctx->ac, t_list, vertex_index, voffset,
channels_per_fetch, false, true); channels_per_fetch, 0, true);
} }
if (num_fetches == 1 && channels_per_fetch > 1) { if (num_fetches == 1 && channels_per_fetch > 1) {
@@ -967,14 +966,14 @@ static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base,
if (swizzle == ~0) { if (swizzle == ~0) {
value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset, value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
0, 1, 0, can_speculate, false); 0, ac_glc, can_speculate, false);
return LLVMBuildBitCast(ctx->ac.builder, value, vec_type, ""); return LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
} }
if (!llvm_type_is_64bit(ctx, type)) { if (!llvm_type_is_64bit(ctx, type)) {
value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset, value = ac_build_buffer_load(&ctx->ac, buffer, 4, NULL, base, offset,
0, 1, 0, can_speculate, false); 0, ac_glc, can_speculate, false);
value = LLVMBuildBitCast(ctx->ac.builder, value, vec_type, ""); value = LLVMBuildBitCast(ctx->ac.builder, value, vec_type, "");
return LLVMBuildExtractElement(ctx->ac.builder, value, return LLVMBuildExtractElement(ctx->ac.builder, value,
@@ -982,10 +981,10 @@ static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base,
} }
value = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset, value = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
swizzle * 4, 1, 0, can_speculate, false); swizzle * 4, ac_glc, can_speculate, false);
value2 = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset, value2 = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
swizzle * 4 + 4, 1, 0, can_speculate, false); swizzle * 4 + 4, ac_glc, can_speculate, false);
return si_llvm_emit_fetch_64bit(bld_base, type, value, value2); return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
} }
@@ -1589,14 +1588,14 @@ LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi,
soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle) * 256, 0); soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle) * 256, 0);
value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0, value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0,
vtx_offset, soffset, 0, 1, 0, true, false); vtx_offset, soffset, 0, ac_glc, true, false);
if (llvm_type_is_64bit(ctx, type)) { if (llvm_type_is_64bit(ctx, type)) {
LLVMValueRef value2; LLVMValueRef value2;
soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle + 1) * 256, 0); soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle + 1) * 256, 0);
value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1,
ctx->i32_0, vtx_offset, soffset, ctx->i32_0, vtx_offset, soffset,
0, 1, 0, true, false); 0, ac_glc, true, false);
return si_llvm_emit_fetch_64bit(bld_base, type, value, value2); return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
} }
return LLVMBuildBitCast(ctx->ac.builder, value, type, ""); return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
@@ -1908,7 +1907,7 @@ static LLVMValueRef buffer_load_const(struct si_shader_context *ctx,
LLVMValueRef offset) LLVMValueRef offset)
{ {
return ac_build_buffer_load(&ctx->ac, resource, 1, NULL, offset, NULL, return ac_build_buffer_load(&ctx->ac, resource, 1, NULL, offset, NULL,
0, 0, 0, true, true); 0, 0, true, true);
} }
static LLVMValueRef load_sample_position(struct ac_shader_abi *abi, LLVMValueRef sample_id) static LLVMValueRef load_sample_position(struct ac_shader_abi *abi, LLVMValueRef sample_id)
@@ -5787,7 +5786,7 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
ac_build_buffer_load(&ctx.ac, ac_build_buffer_load(&ctx.ac,
ctx.gsvs_ring[0], 1, ctx.gsvs_ring[0], 1,
ctx.i32_0, voffset, ctx.i32_0, voffset,
soffset, 0, 1, 1, soffset, 0, ac_glc | ac_slc,
true, false); true, false);
} }
} }

10
src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
View File

@@ -515,7 +515,7 @@ static void load_emit(
emit_data->output[emit_data->chan] = emit_data->output[emit_data->chan] =
ac_build_buffer_load(&ctx->ac, args.resource, ac_build_buffer_load(&ctx->ac, args.resource,
util_last_bit(inst->Dst[0].Register.WriteMask), util_last_bit(inst->Dst[0].Register.WriteMask),
NULL, voffset, NULL, 0, 0, 0, true, true); NULL, voffset, NULL, 0, 0, true, true);
return; return;
} }
@@ -552,9 +552,7 @@ static void load_emit(
ac_build_buffer_load(&ctx->ac, args.resource, ac_build_buffer_load(&ctx->ac, args.resource,
util_last_bit(inst->Dst[0].Register.WriteMask), util_last_bit(inst->Dst[0].Register.WriteMask),
NULL, voffset, NULL, 0, NULL, voffset, NULL, 0,
!!(args.cache_policy & ac_glc), args.cache_policy, can_speculate, false);
!!(args.cache_policy & ac_slc),
can_speculate, false);
return; return;
} }
@@ -566,7 +564,7 @@ static void load_emit(
vindex, vindex,
ctx->i32_0, ctx->i32_0,
num_channels, num_channels,
!!(args.cache_policy & ac_glc), args.cache_policy,
can_speculate); can_speculate);
emit_data->output[emit_data->chan] = emit_data->output[emit_data->chan] =
ac_build_expand_to_vec4(&ctx->ac, result, num_channels); ac_build_expand_to_vec4(&ctx->ac, result, num_channels);
@@ -1340,7 +1338,7 @@ static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
args.resource, args.resource,
vindex, vindex,
ctx->i32_0, ctx->i32_0,
num_channels, false, true); num_channels, 0, true);
emit_data->output[emit_data->chan] = emit_data->output[emit_data->chan] =
ac_build_expand_to_vec4(&ctx->ac, result, num_channels); ac_build_expand_to_vec4(&ctx->ac, result, num_channels);
return; return;