struct si_shader_output_values
{
LLVMValueRef values[4];
- unsigned name;
- unsigned sid;
+ unsigned semantic_name;
+ unsigned semantic_index;
+ ubyte vertex_stream[4];
};
static void si_init_shader_ctx(struct si_shader_context *ctx,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data);
-static void si_dump_shader_key(unsigned shader, union si_shader_key *key,
+static void si_dump_shader_key(unsigned shader, struct si_shader_key *key,
FILE *f);
static void si_build_vs_prolog_function(struct si_shader_context *ctx,
union si_shader_part_key *key);
static void si_build_vs_epilog_function(struct si_shader_context *ctx,
union si_shader_part_key *key);
+static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
+ union si_shader_part_key *key);
static void si_build_ps_prolog_function(struct si_shader_context *ctx,
union si_shader_part_key *key);
static void si_build_ps_epilog_function(struct si_shader_context *ctx,
case TGSI_SEMANTIC_GENERIC:
if (index <= 63-4)
return 4 + index;
- else
- /* same explanation as in the default statement,
- * the only user hitting this is st/nine.
- */
- return 0;
+
+ assert(!"invalid generic index");
+ return 0;
/* patch indices are completely separate and thus start from 0 */
case TGSI_SEMANTIC_TESSOUTER:
return 2 + index;
default:
- /* Don't fail here. The result of this function is only used
- * for LS, TCS, TES, and GS, where legacy GL semantics can't
- * occur, but this function is called for all vertex shaders
- * before it's known whether LS will be compiled or not.
- */
+ assert(!"invalid semantic name");
+ return 0;
+ }
+}
+
+unsigned si_shader_io_get_unique_index2(unsigned name, unsigned index)
+{
+ switch (name) {
+ case TGSI_SEMANTIC_FOG:
+ return 0;
+ case TGSI_SEMANTIC_LAYER:
+ return 1;
+ case TGSI_SEMANTIC_VIEWPORT_INDEX:
+ return 2;
+ case TGSI_SEMANTIC_PRIMID:
+ return 3;
+ case TGSI_SEMANTIC_COLOR: /* these alias */
+ case TGSI_SEMANTIC_BCOLOR:
+ return 4 + index;
+ case TGSI_SEMANTIC_TEXCOORD:
+ return 6 + index;
+ default:
+ assert(!"invalid semantic name");
return 0;
}
}
}
static void declare_input_vs(
- struct si_shader_context *radeon_bld,
+ struct si_shader_context *ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl,
LLVMValueRef out[4])
{
- struct lp_build_context *base = &radeon_bld->soa.bld_base.base;
+ struct lp_build_context *base = &ctx->soa.bld_base.base;
struct gallivm_state *gallivm = base->gallivm;
- struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->soa.bld_base);
- unsigned divisor =
- ctx->shader->key.vs.prolog.instance_divisors[input_index];
unsigned chan;
+ unsigned fix_fetch;
LLVMValueRef t_list_ptr;
LLVMValueRef t_offset;
/* Build the attribute offset */
attribute_offset = lp_build_const_int32(gallivm, 0);
- if (!ctx->no_prolog) {
- buffer_index = LLVMGetParam(radeon_bld->main_fn,
- ctx->param_vertex_index0 +
- input_index);
- } else if (divisor) {
- /* Build index from instance ID, start instance and divisor */
- ctx->shader->info.uses_instanceid = true;
- buffer_index = get_instance_index_for_fetch(ctx,
- SI_PARAM_START_INSTANCE,
- divisor);
- } else {
- /* Load the buffer index for vertices. */
- LLVMValueRef vertex_id = LLVMGetParam(ctx->main_fn,
- ctx->param_vertex_id);
- LLVMValueRef base_vertex = LLVMGetParam(radeon_bld->main_fn,
- SI_PARAM_BASE_VERTEX);
- buffer_index = LLVMBuildAdd(gallivm->builder, base_vertex, vertex_id, "");
- }
+ buffer_index = LLVMGetParam(ctx->main_fn,
+ ctx->param_vertex_index0 +
+ input_index);
args[0] = t_list;
args[1] = attribute_offset;
args[2] = buffer_index;
input = lp_build_intrinsic(gallivm->builder,
"llvm.SI.vs.load.input", ctx->v4f32, args, 3,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
/* Break up the vec4 into individual components */
for (chan = 0; chan < 4; chan++) {
out[chan] = LLVMBuildExtractElement(gallivm->builder,
input, llvm_chan, "");
}
+
+ fix_fetch = (ctx->shader->key.mono.vs.fix_fetch >> (2 * input_index)) & 3;
+ if (fix_fetch) {
+ /* The hardware returns an unsigned value; convert it to a
+ * signed one.
+ */
+ LLVMValueRef tmp = out[3];
+ LLVMValueRef c30 = LLVMConstInt(ctx->i32, 30, 0);
+
+ /* First, recover the sign-extended signed integer value. */
+ if (fix_fetch == SI_FIX_FETCH_A2_SSCALED)
+ tmp = LLVMBuildFPToUI(gallivm->builder, tmp, ctx->i32, "");
+ else
+ tmp = LLVMBuildBitCast(gallivm->builder, tmp, ctx->i32, "");
+
+ /* For the integer-like cases, do a natural sign extension.
+ *
+ * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
+ * and happen to contain 0, 1, 2, 3 as the two LSBs of the
+ * exponent.
+ */
+ tmp = LLVMBuildShl(gallivm->builder, tmp,
+ fix_fetch == SI_FIX_FETCH_A2_SNORM ?
+ LLVMConstInt(ctx->i32, 7, 0) : c30, "");
+ tmp = LLVMBuildAShr(gallivm->builder, tmp, c30, "");
+
+ /* Convert back to the right type. */
+ if (fix_fetch == SI_FIX_FETCH_A2_SNORM) {
+ LLVMValueRef clamp;
+ LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
+ tmp = LLVMBuildSIToFP(gallivm->builder, tmp, ctx->f32, "");
+ clamp = LLVMBuildFCmp(gallivm->builder, LLVMRealULT, tmp, neg_one, "");
+ tmp = LLVMBuildSelect(gallivm->builder, clamp, neg_one, tmp, "");
+ } else if (fix_fetch == SI_FIX_FETCH_A2_SSCALED) {
+ tmp = LLVMBuildSIToFP(gallivm->builder, tmp, ctx->f32, "");
+ }
+
+ out[3] = tmp;
+ }
}
static LLVMValueRef get_primitive_id(struct lp_build_tgsi_context *bld_base,
type_names[func]);
return lp_build_intrinsic(gallivm->builder, name, types[func], args,
- ARRAY_SIZE(args), LLVMReadOnlyAttribute);
+ ARRAY_SIZE(args), LP_FUNC_ATTR_READONLY);
} else {
LLVMValueRef args[] = {
LLVMBuildBitCast(gallivm->builder, rsrc, ctx->v16i8, ""),
type_names[func], arg_type);
return lp_build_intrinsic(gallivm->builder, name, types[func], args,
- ARRAY_SIZE(args), LLVMReadOnlyAttribute);
+ ARRAY_SIZE(args), LP_FUNC_ATTR_READONLY);
}
}
value = lp_build_intrinsic(gallivm->builder,
"llvm.SI.buffer.load.dword.i32.i32",
ctx->i32, args, 9,
- LLVMReadOnlyAttribute);
+ LP_FUNC_ATTR_READONLY);
if (tgsi_type_is_64bit(type)) {
LLVMValueRef value2;
args[2] = lp_build_const_int32(gallivm, (param * 4 + swizzle + 1) * 256);
value2 = lp_build_intrinsic(gallivm->builder,
"llvm.SI.buffer.load.dword.i32.i32",
ctx->i32, args, 9,
- LLVMReadOnlyAttribute);
+ LP_FUNC_ATTR_READONLY);
return si_llvm_emit_fetch_64bit(bld_base, type,
value, value2);
}
}
}
-/* This shouldn't be used by explicit INTERP opcodes. */
-static unsigned select_interp_param(struct si_shader_context *ctx,
- unsigned param)
-{
- if (!ctx->no_prolog)
- return param;
+static LLVMValueRef build_fs_interp(
+ struct lp_build_tgsi_context *bld_base,
+ LLVMValueRef llvm_chan,
+ LLVMValueRef attr_number,
+ LLVMValueRef params,
+ LLVMValueRef i,
+ LLVMValueRef j) {
- if (ctx->shader->key.ps.prolog.force_persp_sample_interp) {
- switch (param) {
- case SI_PARAM_PERSP_CENTROID:
- case SI_PARAM_PERSP_CENTER:
- return SI_PARAM_PERSP_SAMPLE;
- }
- }
- if (ctx->shader->key.ps.prolog.force_linear_sample_interp) {
- switch (param) {
- case SI_PARAM_LINEAR_CENTROID:
- case SI_PARAM_LINEAR_CENTER:
- return SI_PARAM_LINEAR_SAMPLE;
- }
- }
- if (ctx->shader->key.ps.prolog.force_persp_center_interp) {
- switch (param) {
- case SI_PARAM_PERSP_CENTROID:
- case SI_PARAM_PERSP_SAMPLE:
- return SI_PARAM_PERSP_CENTER;
- }
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMValueRef args[5];
+ LLVMValueRef p1;
+ if (HAVE_LLVM < 0x0400) {
+ LLVMValueRef ij[2];
+ ij[0] = LLVMBuildBitCast(gallivm->builder, i, ctx->i32, "");
+ ij[1] = LLVMBuildBitCast(gallivm->builder, j, ctx->i32, "");
+
+ args[0] = llvm_chan;
+ args[1] = attr_number;
+ args[2] = params;
+ args[3] = lp_build_gather_values(gallivm, ij, 2);
+ return lp_build_intrinsic(gallivm->builder, "llvm.SI.fs.interp",
+ ctx->f32, args, 4,
+ LP_FUNC_ATTR_READNONE);
}
- if (ctx->shader->key.ps.prolog.force_linear_center_interp) {
- switch (param) {
- case SI_PARAM_LINEAR_CENTROID:
- case SI_PARAM_LINEAR_SAMPLE:
- return SI_PARAM_LINEAR_CENTER;
- }
+
+ args[0] = i;
+ args[1] = llvm_chan;
+ args[2] = attr_number;
+ args[3] = params;
+
+ p1 = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.interp.p1",
+ ctx->f32, args, 4, LP_FUNC_ATTR_READNONE);
+
+ args[0] = p1;
+ args[1] = j;
+ args[2] = llvm_chan;
+ args[3] = attr_number;
+ args[4] = params;
+
+ return lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.interp.p2",
+ ctx->f32, args, 5, LP_FUNC_ATTR_READNONE);
+}
+
+static LLVMValueRef build_fs_interp_mov(
+ struct lp_build_tgsi_context *bld_base,
+ LLVMValueRef parameter,
+ LLVMValueRef llvm_chan,
+ LLVMValueRef attr_number,
+ LLVMValueRef params) {
+
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMValueRef args[4];
+ if (HAVE_LLVM < 0x0400) {
+ args[0] = llvm_chan;
+ args[1] = attr_number;
+ args[2] = params;
+
+ return lp_build_intrinsic(gallivm->builder,
+ "llvm.SI.fs.constant",
+ ctx->f32, args, 3,
+ LP_FUNC_ATTR_READNONE);
}
- return param;
+ args[0] = parameter;
+ args[1] = llvm_chan;
+ args[2] = attr_number;
+ args[3] = params;
+
+ return lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.interp.mov",
+ ctx->f32, args, 4, LP_FUNC_ATTR_READNONE);
}
/**
LLVMValueRef face,
LLVMValueRef result[4])
{
- struct lp_build_context *base = &ctx->soa.bld_base.base;
- struct lp_build_context *uint = &ctx->soa.bld_base.uint_bld;
+ struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_context *base = &bld_base->base;
+ struct lp_build_context *uint = &bld_base->uint_bld;
struct gallivm_state *gallivm = base->gallivm;
- const char *intr_name;
LLVMValueRef attr_number;
+ LLVMValueRef i, j;
unsigned chan;
- attr_number = lp_build_const_int32(gallivm, input_index);
-
/* fs.constant returns the param from the middle vertex, so it's not
* really useful for flat shading. It's meant to be used for custom
* interpolation (but the intrinsic can't fetch from the other two
* to do the right thing. The only reason we use fs.constant is that
* fs.interp cannot be used on integers, because they can be equal
* to NaN.
+ *
+ * When interp is false we will use fs.constant or for newer llvm,
+ * amdgcn.interp.mov.
*/
- intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
+ bool interp = interp_param != NULL;
+
+ attr_number = lp_build_const_int32(gallivm, input_index);
+
+ if (interp) {
+ interp_param = LLVMBuildBitCast(gallivm->builder, interp_param,
+ LLVMVectorType(ctx->f32, 2), "");
+
+ i = LLVMBuildExtractElement(gallivm->builder, interp_param,
+ uint->zero, "");
+ j = LLVMBuildExtractElement(gallivm->builder, interp_param,
+ uint->one, "");
+ }
if (semantic_name == TGSI_SEMANTIC_COLOR &&
- ctx->shader->key.ps.prolog.color_two_side) {
- LLVMValueRef args[4];
+ ctx->shader->key.part.ps.prolog.color_two_side) {
LLVMValueRef is_face_positive;
LLVMValueRef back_attr_number;
is_face_positive = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
face, uint->zero, "");
- args[2] = prim_mask;
- args[3] = interp_param;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
LLVMValueRef front, back;
- args[0] = llvm_chan;
- args[1] = attr_number;
- front = lp_build_intrinsic(gallivm->builder, intr_name,
- ctx->f32, args, args[3] ? 4 : 3,
- LLVMReadNoneAttribute);
-
- args[1] = back_attr_number;
- back = lp_build_intrinsic(gallivm->builder, intr_name,
- ctx->f32, args, args[3] ? 4 : 3,
- LLVMReadNoneAttribute);
+ if (interp) {
+ front = build_fs_interp(bld_base, llvm_chan,
+ attr_number, prim_mask,
+ i, j);
+ back = build_fs_interp(bld_base, llvm_chan,
+ back_attr_number, prim_mask,
+ i, j);
+ } else {
+ front = build_fs_interp_mov(bld_base,
+ lp_build_const_int32(gallivm, 2), /* P0 */
+ llvm_chan, attr_number, prim_mask);
+ back = build_fs_interp_mov(bld_base,
+ lp_build_const_int32(gallivm, 2), /* P0 */
+ llvm_chan, back_attr_number, prim_mask);
+ }
result[chan] = LLVMBuildSelect(gallivm->builder,
is_face_positive,
"");
}
} else if (semantic_name == TGSI_SEMANTIC_FOG) {
- LLVMValueRef args[4];
-
- args[0] = uint->zero;
- args[1] = attr_number;
- args[2] = prim_mask;
- args[3] = interp_param;
- result[0] = lp_build_intrinsic(gallivm->builder, intr_name,
- ctx->f32, args, args[3] ? 4 : 3,
- LLVMReadNoneAttribute);
+ if (interp) {
+ result[0] = build_fs_interp(bld_base, uint->zero,
+ attr_number, prim_mask, i, j);
+ } else {
+ result[0] = build_fs_interp_mov(bld_base, uint->zero,
+ lp_build_const_int32(gallivm, 2), /* P0 */
+ attr_number, prim_mask);
+ }
result[1] =
result[2] = lp_build_const_float(gallivm, 0.0f);
result[3] = lp_build_const_float(gallivm, 1.0f);
} else {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- LLVMValueRef args[4];
LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
- args[0] = llvm_chan;
- args[1] = attr_number;
- args[2] = prim_mask;
- args[3] = interp_param;
- result[chan] = lp_build_intrinsic(gallivm->builder, intr_name,
- ctx->f32, args, args[3] ? 4 : 3,
- LLVMReadNoneAttribute);
- }
- }
-}
-
-/* LLVMGetParam with bc_optimize resolved. */
-static LLVMValueRef get_interp_param(struct si_shader_context *ctx,
- int interp_param_idx)
-{
- LLVMBuilderRef builder = ctx->gallivm.builder;
- LLVMValueRef main_fn = ctx->main_fn;
- LLVMValueRef param = NULL;
-
- /* Handle PRIM_MASK[31] (bc_optimize). */
- if (ctx->no_prolog &&
- ((ctx->shader->key.ps.prolog.bc_optimize_for_persp &&
- interp_param_idx == SI_PARAM_PERSP_CENTROID) ||
- (ctx->shader->key.ps.prolog.bc_optimize_for_linear &&
- interp_param_idx == SI_PARAM_LINEAR_CENTROID))) {
- /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
- * The hw doesn't compute CENTROID if the whole wave only
- * contains fully-covered quads.
- */
- LLVMValueRef bc_optimize =
- LLVMGetParam(main_fn, SI_PARAM_PRIM_MASK);
- bc_optimize = LLVMBuildLShr(builder,
- bc_optimize,
- LLVMConstInt(ctx->i32, 31, 0), "");
- bc_optimize = LLVMBuildTrunc(builder, bc_optimize, ctx->i1, "");
-
- if (ctx->shader->key.ps.prolog.bc_optimize_for_persp &&
- interp_param_idx == SI_PARAM_PERSP_CENTROID) {
- param = LLVMBuildSelect(builder, bc_optimize,
- LLVMGetParam(main_fn,
- SI_PARAM_PERSP_CENTER),
- LLVMGetParam(main_fn,
- SI_PARAM_PERSP_CENTROID),
- "");
- }
- if (ctx->shader->key.ps.prolog.bc_optimize_for_linear &&
- interp_param_idx == SI_PARAM_LINEAR_CENTROID) {
- param = LLVMBuildSelect(builder, bc_optimize,
- LLVMGetParam(main_fn,
- SI_PARAM_LINEAR_CENTER),
- LLVMGetParam(main_fn,
- SI_PARAM_LINEAR_CENTROID),
- "");
+ if (interp) {
+ result[chan] = build_fs_interp(bld_base,
+ llvm_chan, attr_number, prim_mask, i, j);
+ } else {
+ result[chan] = build_fs_interp_mov(bld_base,
+ lp_build_const_int32(gallivm, 2), /* P0 */
+ llvm_chan, attr_number, prim_mask);
+ }
}
}
-
- if (!param)
- param = LLVMGetParam(main_fn, interp_param_idx);
- return param;
}
static void declare_input_fs(
int interp_param_idx;
/* Get colors from input VGPRs (set by the prolog). */
- if (!ctx->no_prolog &&
- decl->Semantic.Name == TGSI_SEMANTIC_COLOR) {
+ if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR) {
unsigned i = decl->Semantic.Index;
unsigned colors_read = shader->selector->info.colors_read;
unsigned mask = colors_read >> (i * 4);
if (interp_param_idx == -1)
return;
else if (interp_param_idx) {
- interp_param_idx = select_interp_param(ctx,
- interp_param_idx);
- interp_param = get_interp_param(ctx, interp_param_idx);
+ interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
}
if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
decl->Interp.Interpolate == TGSI_INTERPOLATE_COLOR &&
- ctx->shader->key.ps.prolog.flatshade_colors)
+ ctx->shader->key.part.ps.prolog.flatshade_colors)
interp_param = NULL; /* load the constant color */
interp_fs_input(ctx, input_index, decl->Semantic.Name,
if (HAVE_LLVM < 0x0308) {
tid = lp_build_intrinsic(gallivm->builder, "llvm.SI.tid",
- ctx->i32, NULL, 0, LLVMReadNoneAttribute);
+ ctx->i32, NULL, 0, LP_FUNC_ATTR_READNONE);
} else {
LLVMValueRef tid_args[2];
tid_args[0] = lp_build_const_int32(gallivm, 0xffffffff);
tid_args[1] = lp_build_const_int32(gallivm, 0);
tid_args[1] = lp_build_intrinsic(gallivm->builder,
"llvm.amdgcn.mbcnt.lo", ctx->i32,
- tid_args, 2, LLVMReadNoneAttribute);
+ tid_args, 2, LP_FUNC_ATTR_READNONE);
tid = lp_build_intrinsic(gallivm->builder,
"llvm.amdgcn.mbcnt.hi", ctx->i32,
- tid_args, 2, LLVMReadNoneAttribute);
+ tid_args, 2, LP_FUNC_ATTR_READNONE);
}
set_range_metadata(ctx, tid, 0, 64);
return tid;
LLVMValueRef args[2] = {resource, offset};
return lp_build_intrinsic(builder, "llvm.SI.load.const", ctx->f32, args, 2,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
}
static LLVMValueRef load_sample_position(struct si_shader_context *radeon_bld, LLVMValueRef sample_id)
value = lp_build_intrinsic(gallivm->builder,
"llvm.amdgcn.ps.live",
ctx->i1, NULL, 0,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
value = LLVMBuildNot(gallivm->builder, value, "");
value = LLVMBuildSExt(gallivm->builder, value, ctx->i32, "");
break;
args[3] = lp_build_const_int32(base->gallivm, target);
if (ctx->type == PIPE_SHADER_FRAGMENT) {
- const union si_shader_key *key = &ctx->shader->key;
- unsigned col_formats = key->ps.epilog.spi_shader_col_format;
+ const struct si_shader_key *key = &ctx->shader->key;
+ unsigned col_formats = key->part.ps.epilog.spi_shader_col_format;
int cbuf = target - V_008DFC_SQ_EXP_MRT;
assert(cbuf >= 0 && cbuf < 8);
spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
- is_int8 = (key->ps.epilog.color_is_int8 >> cbuf) & 0x1;
+ is_int8 = (key->part.ps.epilog.color_is_int8 >> cbuf) & 0x1;
}
args[4] = uint->zero; /* COMPR flag */
packed = lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.packf16",
ctx->i32, pack_args, 2,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
args[chan + 5] =
LLVMBuildBitCast(base->gallivm->builder,
packed, ctx->f32, "");
struct si_shader_context *ctx = si_shader_context(bld_base);
struct gallivm_state *gallivm = bld_base->base.gallivm;
- if (ctx->shader->key.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
+ if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn,
SI_PARAM_ALPHA_REF);
LLVMValueRef alpha_pass =
lp_build_cmp(&bld_base->base,
- ctx->shader->key.ps.epilog.alpha_func,
+ ctx->shader->key.part.ps.epilog.alpha_func,
alpha, alpha_ref);
LLVMValueRef arg =
lp_build_select(&bld_base->base,
coverage = lp_build_intrinsic(gallivm->builder, "llvm.ctpop.i32",
ctx->i32,
- &coverage, 1, LLVMReadNoneAttribute);
+ &coverage, 1, LP_FUNC_ATTR_READNONE);
coverage = LLVMBuildUIToFP(gallivm->builder, coverage,
ctx->f32, "");
}
}
-/* On SI, the vertex shader is responsible for writing streamout data
- * to buffers. */
-static void si_llvm_emit_streamout(struct si_shader_context *ctx,
- struct si_shader_output_values *outputs,
- unsigned noutput)
+static void emit_streamout_output(struct si_shader_context *ctx,
+ LLVMValueRef const *so_buffers,
+ LLVMValueRef const *so_write_offsets,
+ struct pipe_stream_output *stream_out,
+ struct si_shader_output_values *shader_out)
{
- struct pipe_stream_output_info *so = &ctx->shader->selector->so;
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
- int i, j;
- struct lp_build_if_state if_ctx;
- LLVMValueRef so_buffers[4];
- LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
- SI_PARAM_RW_BUFFERS);
+ unsigned buf_idx = stream_out->output_buffer;
+ unsigned start = stream_out->start_component;
+ unsigned num_comps = stream_out->num_components;
+ LLVMValueRef out[4];
- /* Load the descriptors. */
- for (i = 0; i < 4; ++i) {
- if (ctx->shader->selector->so.stride[i]) {
- LLVMValueRef offset = lp_build_const_int32(gallivm,
- SI_VS_STREAMOUT_BUF0 + i);
+ assert(num_comps && num_comps <= 4);
+ if (!num_comps || num_comps > 4)
+ return;
- so_buffers[i] = build_indexed_load_const(ctx, buf_ptr, offset);
+ /* Load the output as int. */
+ for (int j = 0; j < num_comps; j++) {
+ assert(stream_out->stream == shader_out->vertex_stream[start + j]);
+
+ out[j] = LLVMBuildBitCast(builder,
+ shader_out->values[start + j],
+ ctx->i32, "");
+ }
+
+ /* Pack the output. */
+ LLVMValueRef vdata = NULL;
+
+ switch (num_comps) {
+ case 1: /* as i32 */
+ vdata = out[0];
+ break;
+ case 2: /* as v2i32 */
+ case 3: /* as v4i32 (aligned to 4) */
+ case 4: /* as v4i32 */
+ vdata = LLVMGetUndef(LLVMVectorType(ctx->i32, util_next_power_of_two(num_comps)));
+ for (int j = 0; j < num_comps; j++) {
+ vdata = LLVMBuildInsertElement(builder, vdata, out[j],
+ LLVMConstInt(ctx->i32, j, 0), "");
}
+ break;
}
+ build_tbuffer_store_dwords(ctx, so_buffers[buf_idx],
+ vdata, num_comps,
+ so_write_offsets[buf_idx],
+ LLVMConstInt(ctx->i32, 0, 0),
+ stream_out->dst_offset * 4);
+}
+
+/**
+ * Write streamout data to buffers for vertex stream @p stream (different
+ * vertex streams can occur for GS copy shaders).
+ */
+static void si_llvm_emit_streamout(struct si_shader_context *ctx,
+ struct si_shader_output_values *outputs,
+ unsigned noutput, unsigned stream)
+{
+ struct si_shader_selector *sel = ctx->shader->selector;
+ struct pipe_stream_output_info *so = &sel->so;
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ int i;
+ struct lp_build_if_state if_ctx;
+
/* Get bits [22:16], i.e. (so_param >> 16) & 127; */
LLVMValueRef so_vtx_count =
unpack_param(ctx, ctx->param_streamout_config, 16, 7);
LLVMValueRef can_emit =
LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
- LLVMValueRef stream_id =
- unpack_param(ctx, ctx->param_streamout_config, 24, 2);
-
/* Emit the streamout code conditionally. This actually avoids
* out-of-bounds buffer access. The hw tells us via the SGPR
* (so_vtx_count) which threads are allowed to emit streamout data. */
/* Compute (streamout_write_index + thread_id). */
so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");
- /* Compute the write offset for each enabled buffer. */
+ /* Load the descriptor and compute the write offset for each
+ * enabled buffer. */
LLVMValueRef so_write_offset[4] = {};
+ LLVMValueRef so_buffers[4];
+ LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
+ SI_PARAM_RW_BUFFERS);
+
for (i = 0; i < 4; i++) {
if (!so->stride[i])
continue;
+ LLVMValueRef offset = lp_build_const_int32(gallivm,
+ SI_VS_STREAMOUT_BUF0 + i);
+
+ so_buffers[i] = build_indexed_load_const(ctx, buf_ptr, offset);
+
LLVMValueRef so_offset = LLVMGetParam(ctx->main_fn,
ctx->param_streamout_offset[i]);
so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");
/* Write streamout data. */
for (i = 0; i < so->num_outputs; i++) {
- unsigned buf_idx = so->output[i].output_buffer;
unsigned reg = so->output[i].register_index;
- unsigned start = so->output[i].start_component;
- unsigned num_comps = so->output[i].num_components;
- unsigned stream = so->output[i].stream;
- LLVMValueRef out[4];
- struct lp_build_if_state if_ctx_stream;
-
- assert(num_comps && num_comps <= 4);
- if (!num_comps || num_comps > 4)
- continue;
if (reg >= noutput)
continue;
- /* Load the output as int. */
- for (j = 0; j < num_comps; j++) {
- out[j] = LLVMBuildBitCast(builder,
- outputs[reg].values[start+j],
- ctx->i32, "");
- }
-
- /* Pack the output. */
- LLVMValueRef vdata = NULL;
-
- switch (num_comps) {
- case 1: /* as i32 */
- vdata = out[0];
- break;
- case 2: /* as v2i32 */
- case 3: /* as v4i32 (aligned to 4) */
- case 4: /* as v4i32 */
- vdata = LLVMGetUndef(LLVMVectorType(ctx->i32, util_next_power_of_two(num_comps)));
- for (j = 0; j < num_comps; j++) {
- vdata = LLVMBuildInsertElement(builder, vdata, out[j],
- LLVMConstInt(ctx->i32, j, 0), "");
- }
- break;
- }
+ if (stream != so->output[i].stream)
+ continue;
- LLVMValueRef can_emit_stream =
- LLVMBuildICmp(builder, LLVMIntEQ,
- stream_id,
- lp_build_const_int32(gallivm, stream), "");
-
- lp_build_if(&if_ctx_stream, gallivm, can_emit_stream);
- build_tbuffer_store_dwords(ctx, so_buffers[buf_idx],
- vdata, num_comps,
- so_write_offset[buf_idx],
- LLVMConstInt(ctx->i32, 0, 0),
- so->output[i].dst_offset*4);
- lp_build_endif(&if_ctx_stream);
+ emit_streamout_output(ctx, so_buffers, so_write_offset,
+ &so->output[i], &outputs[reg]);
}
}
lp_build_endif(&if_ctx);
unsigned pos_idx;
int i;
- if (outputs && ctx->shader->selector->so.num_outputs) {
- si_llvm_emit_streamout(ctx, outputs, noutput);
- }
-
for (i = 0; i < noutput; i++) {
- semantic_name = outputs[i].name;
- semantic_index = outputs[i].sid;
+ semantic_name = outputs[i].semantic_name;
+ semantic_index = outputs[i].semantic_index;
+ bool export_param = true;
+
+ switch (semantic_name) {
+ case TGSI_SEMANTIC_POSITION: /* ignore these */
+ case TGSI_SEMANTIC_PSIZE:
+ case TGSI_SEMANTIC_CLIPVERTEX:
+ case TGSI_SEMANTIC_EDGEFLAG:
+ break;
+ case TGSI_SEMANTIC_GENERIC:
+ case TGSI_SEMANTIC_CLIPDIST:
+ if (shader->key.opt.hw_vs.kill_outputs &
+ (1ull << si_shader_io_get_unique_index(semantic_name, semantic_index)))
+ export_param = false;
+ break;
+ default:
+ if (shader->key.opt.hw_vs.kill_outputs2 &
+ (1u << si_shader_io_get_unique_index2(semantic_name, semantic_index)))
+ export_param = false;
+ break;
+ }
+
+ if (outputs[i].vertex_stream[0] != 0 &&
+ outputs[i].vertex_stream[1] != 0 &&
+ outputs[i].vertex_stream[2] != 0 &&
+ outputs[i].vertex_stream[3] != 0)
+ export_param = false;
handle_semantic:
/* Select the correct target */
break;
case TGSI_SEMANTIC_COLOR:
case TGSI_SEMANTIC_BCOLOR:
+ if (!export_param)
+ continue;
target = V_008DFC_SQ_EXP_PARAM + param_count;
assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
shader->info.vs_output_param_offset[i] = param_count;
param_count++;
break;
case TGSI_SEMANTIC_CLIPDIST:
+ if (shader->key.opt.hw_vs.clip_disable) {
+ semantic_name = TGSI_SEMANTIC_GENERIC;
+ goto handle_semantic;
+ }
target = V_008DFC_SQ_EXP_POS + 2 + semantic_index;
break;
case TGSI_SEMANTIC_CLIPVERTEX:
+ if (shader->key.opt.hw_vs.clip_disable)
+ continue;
si_llvm_emit_clipvertex(bld_base, pos_args, outputs[i].values);
continue;
case TGSI_SEMANTIC_PRIMID:
case TGSI_SEMANTIC_FOG:
case TGSI_SEMANTIC_TEXCOORD:
case TGSI_SEMANTIC_GENERIC:
+ if (!export_param)
+ continue;
target = V_008DFC_SQ_EXP_PARAM + param_count;
assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
shader->info.vs_output_param_offset[i] = param_count;
}
}
+/**
+ * Forward all outputs from the vertex shader to the TES. This is only used
+ * for the fixed function TCS.
+ */
static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
lds_base = get_tcs_in_current_patch_offset(ctx);
lds_base = LLVMBuildAdd(gallivm->builder, lds_base, lds_vertex_offset, "");
- inputs = ctx->shader->key.tcs.epilog.inputs_to_copy;
+ inputs = ctx->shader->key.mono.tcs.inputs_to_copy;
while (inputs) {
unsigned i = u_bit_scan64(&inputs);
invocation_id, bld_base->uint_bld.zero, ""));
/* Determine the layout of one tess factor element in the buffer. */
- switch (shader->key.tcs.epilog.prim_mode) {
+ switch (shader->key.part.tcs.epilog.prim_mode) {
case PIPE_PRIM_LINES:
stride = 2; /* 2 dwords, 1 vec2 store */
outer_comps = 2;
lp_build_const_int32(gallivm,
tess_outer_index * 4), "");
- for (i = 0; i < outer_comps; i++)
- out[i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_outer);
- for (i = 0; i < inner_comps; i++)
- out[outer_comps+i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_inner);
+ if (shader->key.part.tcs.epilog.prim_mode == PIPE_PRIM_LINES) {
+ /* For isolines, the hardware expects tess factors in the
+ * reverse order from what GLSL / TGSI specify.
+ */
+ out[0] = lds_load(bld_base, TGSI_TYPE_SIGNED, 1, lds_outer);
+ out[1] = lds_load(bld_base, TGSI_TYPE_SIGNED, 0, lds_outer);
+ } else {
+ for (i = 0; i < outer_comps; i++)
+ out[i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_outer);
+ for (i = 0; i < inner_comps; i++)
+ out[outer_comps+i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_inner);
+ }
/* Convert the outputs to vectors for stores. */
vec0 = lp_build_gather_values(gallivm, out, MIN2(stride, 4));
struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef rel_patch_id, invocation_id, tf_lds_offset;
+ si_copy_tcs_inputs(bld_base);
+
rel_patch_id = get_rel_patch_id(ctx);
invocation_id = unpack_param(ctx, SI_PARAM_REL_IDS, 8, 5);
tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);
- if (!ctx->no_epilog) {
- /* Return epilog parameters from this function. */
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- LLVMValueRef ret = ctx->return_value;
- LLVMValueRef rw_buffers, rw0, rw1, tf_soffset;
- unsigned vgpr;
-
- /* RW_BUFFERS pointer */
- rw_buffers = LLVMGetParam(ctx->main_fn,
- SI_PARAM_RW_BUFFERS);
- rw_buffers = LLVMBuildPtrToInt(builder, rw_buffers, ctx->i64, "");
- rw_buffers = LLVMBuildBitCast(builder, rw_buffers, ctx->v2i32, "");
- rw0 = LLVMBuildExtractElement(builder, rw_buffers,
- bld_base->uint_bld.zero, "");
- rw1 = LLVMBuildExtractElement(builder, rw_buffers,
- bld_base->uint_bld.one, "");
- ret = LLVMBuildInsertValue(builder, ret, rw0, 0, "");
- ret = LLVMBuildInsertValue(builder, ret, rw1, 1, "");
-
- /* Tess factor buffer soffset is after user SGPRs. */
- tf_soffset = LLVMGetParam(ctx->main_fn,
- SI_PARAM_TESS_FACTOR_OFFSET);
- ret = LLVMBuildInsertValue(builder, ret, tf_soffset,
- SI_TCS_NUM_USER_SGPR + 1, "");
-
- /* VGPRs */
- rel_patch_id = bitcast(bld_base, TGSI_TYPE_FLOAT, rel_patch_id);
- invocation_id = bitcast(bld_base, TGSI_TYPE_FLOAT, invocation_id);
- tf_lds_offset = bitcast(bld_base, TGSI_TYPE_FLOAT, tf_lds_offset);
-
- vgpr = SI_TCS_NUM_USER_SGPR + 2;
- ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, "");
- ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, "");
- ret = LLVMBuildInsertValue(builder, ret, tf_lds_offset, vgpr++, "");
- ctx->return_value = ret;
- return;
- }
+ /* Return epilog parameters from this function. */
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMValueRef ret = ctx->return_value;
+ LLVMValueRef rw_buffers, rw0, rw1, tf_soffset;
+ unsigned vgpr;
- si_copy_tcs_inputs(bld_base);
- si_write_tess_factors(bld_base, rel_patch_id, invocation_id, tf_lds_offset);
+ /* RW_BUFFERS pointer */
+ rw_buffers = LLVMGetParam(ctx->main_fn,
+ SI_PARAM_RW_BUFFERS);
+ rw_buffers = LLVMBuildPtrToInt(builder, rw_buffers, ctx->i64, "");
+ rw_buffers = LLVMBuildBitCast(builder, rw_buffers, ctx->v2i32, "");
+ rw0 = LLVMBuildExtractElement(builder, rw_buffers,
+ bld_base->uint_bld.zero, "");
+ rw1 = LLVMBuildExtractElement(builder, rw_buffers,
+ bld_base->uint_bld.one, "");
+ ret = LLVMBuildInsertValue(builder, ret, rw0, 0, "");
+ ret = LLVMBuildInsertValue(builder, ret, rw1, 1, "");
+
+ /* Tess factor buffer soffset is after user SGPRs. */
+ tf_soffset = LLVMGetParam(ctx->main_fn,
+ SI_PARAM_TESS_FACTOR_OFFSET);
+ ret = LLVMBuildInsertValue(builder, ret, tf_soffset,
+ SI_TCS_NUM_USER_SGPR + 1, "");
+
+ /* VGPRs */
+ rel_patch_id = bitcast(bld_base, TGSI_TYPE_FLOAT, rel_patch_id);
+ invocation_id = bitcast(bld_base, TGSI_TYPE_FLOAT, invocation_id);
+ tf_lds_offset = bitcast(bld_base, TGSI_TYPE_FLOAT, tf_lds_offset);
+
+ vgpr = SI_TCS_NUM_USER_SGPR + 2;
+ ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, "");
+ ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, "");
+ ret = LLVMBuildInsertValue(builder, ret, tf_lds_offset, vgpr++, "");
+ ctx->return_value = ret;
}
static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context *bld_base)
struct si_shader_output_values *outputs = NULL;
int i,j;
- assert(!ctx->is_gs_copy_shader);
+ assert(!ctx->shader->is_gs_copy_shader);
outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
}
for (i = 0; i < info->num_outputs; i++) {
- outputs[i].name = info->output_semantic_name[i];
- outputs[i].sid = info->output_semantic_index[i];
+ outputs[i].semantic_name = info->output_semantic_name[i];
+ outputs[i].semantic_index = info->output_semantic_index[i];
- for (j = 0; j < 4; j++)
+ for (j = 0; j < 4; j++) {
outputs[i].values[j] =
LLVMBuildLoad(gallivm->builder,
ctx->soa.outputs[i][j],
"");
- }
-
- if (ctx->no_epilog) {
- /* Export PrimitiveID when PS needs it. */
- if (si_vs_exports_prim_id(ctx->shader)) {
- outputs[i].name = TGSI_SEMANTIC_PRIMID;
- outputs[i].sid = 0;
- outputs[i].values[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- get_primitive_id(bld_base, 0));
- outputs[i].values[1] = bld_base->base.undef;
- outputs[i].values[2] = bld_base->base.undef;
- outputs[i].values[3] = bld_base->base.undef;
- i++;
+ outputs[i].vertex_stream[j] =
+ (info->output_streams[i] >> (2 * j)) & 3;
}
- } else {
- /* Return the primitive ID from the LLVM function. */
- ctx->return_value =
- LLVMBuildInsertValue(gallivm->builder,
- ctx->return_value,
- bitcast(bld_base, TGSI_TYPE_FLOAT,
- get_primitive_id(bld_base, 0)),
- VS_EPILOG_PRIMID_LOC, "");
+
}
+ /* Return the primitive ID from the LLVM function. */
+ ctx->return_value =
+ LLVMBuildInsertValue(gallivm->builder,
+ ctx->return_value,
+ bitcast(bld_base, TGSI_TYPE_FLOAT,
+ get_primitive_id(bld_base, 0)),
+ VS_EPILOG_PRIMID_LOC, "");
+
+ if (ctx->shader->selector->so.num_outputs)
+ si_llvm_emit_streamout(ctx, outputs, i, 0);
si_llvm_export_vs(bld_base, outputs, i);
FREE(outputs);
}
}
}
- /* SI (except OLAND) has a bug that it only looks
+ /* SI (except OLAND and HAINAN) has a bug that it only looks
* at the X writemask component. */
if (ctx->screen->b.chip_class == SI &&
- ctx->screen->b.family != CHIP_OLAND)
+ ctx->screen->b.family != CHIP_OLAND &&
+ ctx->screen->b.family != CHIP_HAINAN)
mask |= 0x1;
/* Specify which components to enable */
int i;
/* Clamp color */
- if (ctx->shader->key.ps.epilog.clamp_color)
+ if (ctx->shader->key.part.ps.epilog.clamp_color)
for (i = 0; i < 4; i++)
color[i] = si_llvm_saturate(bld_base, color[i]);
/* Alpha to one */
- if (ctx->shader->key.ps.epilog.alpha_to_one)
+ if (ctx->shader->key.part.ps.epilog.alpha_to_one)
color[3] = base->one;
/* Alpha test */
if (index == 0 &&
- ctx->shader->key.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
+ ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
si_alpha_test(bld_base, color[3]);
/* Line & polygon smoothing */
- if (ctx->shader->key.ps.epilog.poly_line_smoothing)
+ if (ctx->shader->key.part.ps.epilog.poly_line_smoothing)
color[3] = si_scale_alpha_by_sample_mask(bld_base, color[3],
samplemask_param);
/* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
- if (ctx->shader->key.ps.epilog.last_cbuf > 0) {
+ if (ctx->shader->key.part.ps.epilog.last_cbuf > 0) {
LLVMValueRef args[8][9];
int c, last = -1;
/* Get the export arguments, also find out what the last one is. */
- for (c = 0; c <= ctx->shader->key.ps.epilog.last_cbuf; c++) {
+ for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
si_llvm_init_export_args(bld_base, color,
V_008DFC_SQ_EXP_MRT + c, args[c]);
if (args[c][0] != bld_base->uint_bld.zero)
}
/* Emit all exports. */
- for (c = 0; c <= ctx->shader->key.ps.epilog.last_cbuf; c++) {
+ for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
if (is_last && last == c) {
args[c][1] = bld_base->uint_bld.one; /* whether the EXEC mask is valid */
args[c][2] = bld_base->uint_bld.one; /* DONE bit */
ctx->voidt, args, 9, 0);
}
-static void si_llvm_emit_fs_epilogue(struct lp_build_tgsi_context *bld_base)
-{
- struct si_shader_context *ctx = si_shader_context(bld_base);
- struct si_shader *shader = ctx->shader;
- struct lp_build_context *base = &bld_base->base;
- struct tgsi_shader_info *info = &shader->selector->info;
- LLVMBuilderRef builder = base->gallivm->builder;
- LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
- int last_color_export = -1;
- int i;
- struct si_ps_exports exp = {};
-
- /* Determine the last export. If MRTZ is present, it's always last.
- * Otherwise, find the last color export.
- */
- if (!info->writes_z && !info->writes_stencil && !info->writes_samplemask) {
- unsigned spi_format = shader->key.ps.epilog.spi_shader_col_format;
-
- /* Don't export NULL and return if alpha-test is enabled. */
- if (shader->key.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS &&
- shader->key.ps.epilog.alpha_func != PIPE_FUNC_NEVER &&
- (spi_format & 0xf) == 0)
- spi_format |= V_028714_SPI_SHADER_32_AR;
-
- for (i = 0; i < info->num_outputs; i++) {
- unsigned index = info->output_semantic_index[i];
-
- if (info->output_semantic_name[i] != TGSI_SEMANTIC_COLOR)
- continue;
-
- /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
- if (shader->key.ps.epilog.last_cbuf > 0) {
- /* Just set this if any of the colorbuffers are enabled. */
- if (spi_format &
- ((1llu << (4 * (shader->key.ps.epilog.last_cbuf + 1))) - 1))
- last_color_export = i;
- continue;
- }
-
- if ((spi_format >> (index * 4)) & 0xf)
- last_color_export = i;
- }
-
- /* If there are no outputs, export NULL. */
- if (last_color_export == -1) {
- si_export_null(bld_base);
- return;
- }
- }
-
- for (i = 0; i < info->num_outputs; i++) {
- unsigned semantic_name = info->output_semantic_name[i];
- unsigned semantic_index = info->output_semantic_index[i];
- unsigned j;
- LLVMValueRef color[4] = {};
-
- /* Select the correct target */
- switch (semantic_name) {
- case TGSI_SEMANTIC_POSITION:
- depth = LLVMBuildLoad(builder,
- ctx->soa.outputs[i][2], "");
- break;
- case TGSI_SEMANTIC_STENCIL:
- stencil = LLVMBuildLoad(builder,
- ctx->soa.outputs[i][1], "");
- break;
- case TGSI_SEMANTIC_SAMPLEMASK:
- samplemask = LLVMBuildLoad(builder,
- ctx->soa.outputs[i][0], "");
- break;
- case TGSI_SEMANTIC_COLOR:
- for (j = 0; j < 4; j++)
- color[j] = LLVMBuildLoad(builder,
- ctx->soa.outputs[i][j], "");
-
- si_export_mrt_color(bld_base, color, semantic_index,
- SI_PARAM_SAMPLE_COVERAGE,
- last_color_export == i, &exp);
- break;
- default:
- fprintf(stderr,
- "Warning: SI unhandled fs output type:%d\n",
- semantic_name);
- }
- }
-
- if (depth || stencil || samplemask)
- si_export_mrt_z(bld_base, depth, stencil, samplemask, &exp);
-
- si_emit_ps_exports(ctx, &exp);
-}
-
/**
* Return PS outputs in this order:
*
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef size =
LLVMBuildExtractElement(builder, descriptor,
- lp_build_const_int32(gallivm, 6), "");
+ lp_build_const_int32(gallivm, 2), "");
if (ctx->screen->b.chip_class >= VI) {
/* On VI, the descriptor contains the size in bytes,
*/
LLVMValueRef stride =
LLVMBuildExtractElement(builder, descriptor,
- lp_build_const_int32(gallivm, 5), "");
+ lp_build_const_int32(gallivm, 1), "");
stride = LLVMBuildLShr(builder, stride,
lp_build_const_int32(gallivm, 16), "");
stride = LLVMBuildAnd(builder, stride,
* point in the program by emitting empty inline assembly that is marked as
* having side effects.
*/
+#if 0 /* unused currently */
static void emit_optimization_barrier(struct si_shader_context *ctx)
{
LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, "", "", true, false);
LLVMBuildCall(builder, inlineasm, NULL, 0, "");
}
+#endif
-static void emit_waitcnt(struct si_shader_context *ctx)
+/* Combine these with & instead of |. */
+#define NOOP_WAITCNT 0xf7f
+#define LGKM_CNT 0x07f
+#define VM_CNT 0xf70
+
+static void emit_waitcnt(struct si_shader_context *ctx, unsigned simm16)
{
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef args[1] = {
- lp_build_const_int32(gallivm, 0xf70)
+ lp_build_const_int32(gallivm, simm16)
};
lp_build_intrinsic(builder, "llvm.amdgcn.s.waitcnt",
ctx->voidt, args, 1, 0);
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef src0 = lp_build_emit_fetch(bld_base, emit_data->inst, 0, 0);
+ unsigned flags = LLVMConstIntGetZExtValue(src0);
+ unsigned waitcnt = NOOP_WAITCNT;
+
+ if (flags & TGSI_MEMBAR_THREAD_GROUP)
+ waitcnt &= VM_CNT & LGKM_CNT;
- emit_waitcnt(ctx);
+ if (flags & (TGSI_MEMBAR_ATOMIC_BUFFER |
+ TGSI_MEMBAR_SHADER_BUFFER |
+ TGSI_MEMBAR_SHADER_IMAGE))
+ waitcnt &= VM_CNT;
+
+ if (flags & TGSI_MEMBAR_SHARED)
+ waitcnt &= LGKM_CNT;
+
+ if (waitcnt != NOOP_WAITCNT)
+ emit_waitcnt(ctx, waitcnt);
}
static LLVMValueRef
}
}
-/**
- * Load the resource descriptor for \p image.
+static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
+{
+ return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
+ CONST_ADDR_SPACE);
+}
+
+/**
+ * Load the resource descriptor for \p image.
*/
static void
image_fetch_rsrc(
struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_src_register *image,
- bool dcc_off,
+ bool is_store, unsigned target,
LLVMValueRef *rsrc)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
SI_PARAM_IMAGES);
LLVMValueRef index, tmp;
+ bool dcc_off = target != TGSI_TEXTURE_BUFFER && is_store;
assert(image->Register.File == TGSI_FILE_IMAGE);
index = LLVMConstInt(ctx->i32, image->Register.Index, 0);
if (info->images_writemask & (1 << image->Register.Index) &&
- !(info->images_buffers & (1 << image->Register.Index)))
+ target != TGSI_TEXTURE_BUFFER)
dcc_off = true;
} else {
/* From the GL_ARB_shader_image_load_store extension spec:
SI_NUM_IMAGES);
}
+ if (target == TGSI_TEXTURE_BUFFER) {
+ LLVMBuilderRef builder = ctx->gallivm.builder;
+
+ rsrc_ptr = LLVMBuildPointerCast(builder, rsrc_ptr,
+ const_array(ctx->v4i32, 0), "");
+ index = LLVMBuildMul(builder, index,
+ LLVMConstInt(ctx->i32, 2, 0), "");
+ index = LLVMBuildAdd(builder, index,
+ LLVMConstInt(ctx->i32, 1, 0), "");
+ *rsrc = build_indexed_load_const(ctx, rsrc_ptr, index);
+ return;
+ }
+
tmp = build_indexed_load_const(ctx, rsrc_ptr, index);
if (dcc_off)
tmp = force_dcc_off(ctx, tmp);
struct si_shader_context *ctx,
struct lp_build_emit_data * emit_data,
unsigned target,
- bool atomic)
+ bool atomic,
+ bool force_glc)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
LLVMValueRef r128 = i1false;
LLVMValueRef da = tgsi_is_array_image(target) ? i1true : i1false;
LLVMValueRef glc =
+ force_glc ||
inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE) ?
i1true : i1false;
LLVMValueRef slc = i1false;
emit_data->args[emit_data->arg_count++] = da;
}
-/**
- * Given a 256 bit resource, extract the top half (which stores the buffer
- * resource in the case of textures and images).
- */
-static LLVMValueRef extract_rsrc_top_half(
- struct si_shader_context *ctx,
- LLVMValueRef rsrc)
-{
- struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
- LLVMTypeRef v2i128 = LLVMVectorType(ctx->i128, 2);
-
- rsrc = LLVMBuildBitCast(gallivm->builder, rsrc, v2i128, "");
- rsrc = LLVMBuildExtractElement(gallivm->builder, rsrc, bld_base->uint_bld.one, "");
- rsrc = LLVMBuildBitCast(gallivm->builder, rsrc, ctx->v4i32, "");
-
- return rsrc;
-}
-
/**
* Append the resource and indexing arguments for buffer intrinsics.
*
LLVMValueRef rsrc,
LLVMValueRef index,
LLVMValueRef offset,
- bool atomic)
+ bool atomic,
+ bool force_glc)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
LLVMValueRef i1false = LLVMConstInt(ctx->i1, 0, 0);
emit_data->args[emit_data->arg_count++] = offset; /* voffset */
if (!atomic) {
emit_data->args[emit_data->arg_count++] =
+ force_glc ||
inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE) ?
i1true : i1false; /* glc */
}
offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
- offset, false);
+ offset, false, false);
} else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
LLVMValueRef coords;
- image_fetch_rsrc(bld_base, &inst->Src[0], false, &rsrc);
+ image_fetch_rsrc(bld_base, &inst->Src[0], false, target, &rsrc);
coords = image_fetch_coords(bld_base, inst, 1);
if (target == TGSI_TEXTURE_BUFFER) {
- rsrc = extract_rsrc_top_half(ctx, rsrc);
buffer_append_args(ctx, emit_data, rsrc, coords,
- bld_base->uint_bld.zero, false);
+ bld_base->uint_bld.zero, false, false);
} else {
emit_data->args[0] = coords;
emit_data->args[1] = rsrc;
emit_data->args[2] = lp_build_const_int32(gallivm, 15); /* dmask */
emit_data->arg_count = 3;
- image_append_args(ctx, emit_data, target, false);
+ image_append_args(ctx, emit_data, target, false, false);
}
}
}
emit_data->output[emit_data->chan] = lp_build_intrinsic(
builder, intrinsic_name, dst_type,
emit_data->args, emit_data->arg_count,
- LLVMReadOnlyAttribute);
+ LP_FUNC_ATTR_READONLY);
}
static LLVMValueRef get_memory_ptr(struct si_shader_context *ctx,
}
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
- emit_waitcnt(ctx);
+ emit_waitcnt(ctx, VM_CNT);
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
load_emit_buffer(ctx, emit_data);
lp_build_intrinsic(
builder, "llvm.amdgcn.buffer.load.format.v4f32", emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LLVMReadOnlyAttribute);
+ LP_FUNC_ATTR_READONLY);
} else {
get_image_intr_name("llvm.amdgcn.image.load",
emit_data->dst_type, /* vdata */
lp_build_intrinsic(
builder, intrinsic_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LLVMReadOnlyAttribute);
+ LP_FUNC_ATTR_READONLY);
}
}
offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
- offset, false);
+ offset, false, false);
} else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE) {
unsigned target = inst->Memory.Texture;
LLVMValueRef coords;
+ /* 8bit/16bit TC L1 write corruption bug on SI.
+ * All store opcodes not aligned to a dword are affected.
+ *
+ * The only way to get unaligned stores in radeonsi is through
+ * shader images.
+ */
+ bool force_glc = ctx->screen->b.chip_class == SI;
+
coords = image_fetch_coords(bld_base, inst, 0);
if (target == TGSI_TEXTURE_BUFFER) {
- image_fetch_rsrc(bld_base, &memory, false, &rsrc);
-
- rsrc = extract_rsrc_top_half(ctx, rsrc);
+ image_fetch_rsrc(bld_base, &memory, true, target, &rsrc);
buffer_append_args(ctx, emit_data, rsrc, coords,
- bld_base->uint_bld.zero, false);
+ bld_base->uint_bld.zero, false, force_glc);
} else {
emit_data->args[1] = coords;
- image_fetch_rsrc(bld_base, &memory, true, &emit_data->args[2]);
+ image_fetch_rsrc(bld_base, &memory, true, target,
+ &emit_data->args[2]);
emit_data->args[3] = lp_build_const_int32(gallivm, 15); /* dmask */
emit_data->arg_count = 4;
- image_append_args(ctx, emit_data, target, false);
+ image_append_args(ctx, emit_data, target, false, force_glc);
}
}
}
}
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
- emit_waitcnt(ctx);
+ emit_waitcnt(ctx, VM_CNT);
if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
store_emit_buffer(ctx, emit_data);
offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
- offset, true);
+ offset, true, false);
} else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
unsigned target = inst->Memory.Texture;
LLVMValueRef coords;
- image_fetch_rsrc(bld_base, &inst->Src[0],
- target != TGSI_TEXTURE_BUFFER, &rsrc);
+ image_fetch_rsrc(bld_base, &inst->Src[0], true, target, &rsrc);
coords = image_fetch_coords(bld_base, inst, 1);
if (target == TGSI_TEXTURE_BUFFER) {
- rsrc = extract_rsrc_top_half(ctx, rsrc);
buffer_append_args(ctx, emit_data, rsrc, coords,
- bld_base->uint_bld.zero, true);
+ bld_base->uint_bld.zero, true, false);
} else {
emit_data->args[emit_data->arg_count++] = coords;
emit_data->args[emit_data->arg_count++] = rsrc;
- image_append_args(ctx, emit_data, target, true);
+ image_append_args(ctx, emit_data, target, true, false);
}
}
}
emit_data->args[0] = shader_buffer_fetch_rsrc(ctx, reg);
emit_data->arg_count = 1;
} else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
- image_fetch_rsrc(bld_base, reg, false, &emit_data->args[0]);
+ image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
+ &emit_data->args[0]);
emit_data->arg_count = 1;
} else {
emit_data->args[0] = bld_base->uint_bld.zero; /* mip level */
- image_fetch_rsrc(bld_base, reg, false, &emit_data->args[1]);
+ image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
+ &emit_data->args[1]);
emit_data->args[2] = lp_build_const_int32(gallivm, 15); /* dmask */
emit_data->args[3] = bld_base->uint_bld.zero; /* unorm */
emit_data->args[4] = bld_base->uint_bld.zero; /* r128 */
out = lp_build_intrinsic(
builder, "llvm.SI.getresinfo.i32", emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
/* Divide the number of layers by 6 to get the number of cubes. */
if (inst->Memory.Texture == TGSI_TEXTURE_CUBE_ARRAY) {
enum desc_type {
DESC_IMAGE,
+ DESC_BUFFER,
DESC_FMASK,
- DESC_SAMPLER
+ DESC_SAMPLER,
};
-static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
-{
- return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
- CONST_ADDR_SPACE);
-}
-
/**
* Load an image view, fmask view. or sampler state descriptor.
*/
/* The image is at [0:7]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
break;
+ case DESC_BUFFER:
+ /* The buffer is in [4:7]. */
+ index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
+ index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 1, 0), "");
+ list = LLVMBuildPointerCast(builder, list,
+ const_array(ctx->v4i32, 0), "");
+ break;
case DESC_FMASK:
/* The FMASK is at [8:15]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
index = LLVMConstInt(ctx->i32, sampler_index, 0);
}
- *res_ptr = load_sampler_desc(ctx, index, DESC_IMAGE);
+ if (target == TGSI_TEXTURE_BUFFER)
+ *res_ptr = load_sampler_desc(ctx, index, DESC_BUFFER);
+ else
+ *res_ptr = load_sampler_desc(ctx, index, DESC_IMAGE);
+
+ if (samp_ptr)
+ *samp_ptr = NULL;
+ if (fmask_ptr)
+ *fmask_ptr = NULL;
if (target == TGSI_TEXTURE_2D_MSAA ||
target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
- if (samp_ptr)
- *samp_ptr = NULL;
if (fmask_ptr)
*fmask_ptr = load_sampler_desc(ctx, index, DESC_FMASK);
- } else {
+ } else if (target != TGSI_TEXTURE_BUFFER) {
if (samp_ptr) {
*samp_ptr = load_sampler_desc(ctx, index, DESC_SAMPLER);
*samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
}
- if (fmask_ptr)
- *fmask_ptr = NULL;
}
}
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction *inst = emit_data->inst;
unsigned target = inst->Texture.Texture;
LLVMValueRef res_ptr;
if (target == TGSI_TEXTURE_BUFFER) {
/* Read the size from the buffer descriptor directly. */
- LLVMValueRef res = LLVMBuildBitCast(builder, res_ptr, ctx->v8i32, "");
- emit_data->args[0] = get_buffer_size(bld_base, res);
+ emit_data->args[0] = get_buffer_size(bld_base, res_ptr);
return;
}
emit_data->output[emit_data->chan] = lp_build_intrinsic(
base->gallivm->builder, "llvm.SI.getresinfo.i32",
emit_data->dst_type, emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
/* Divide the number of layers by 6 to get the number of cubes. */
if (target == TGSI_TEXTURE_CUBE_ARRAY ||
tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
if (target == TGSI_TEXTURE_BUFFER) {
- LLVMTypeRef v2i128 = LLVMVectorType(ctx->i128, 2);
-
- /* Bitcast and truncate v8i32 to v16i8. */
- LLVMValueRef res = res_ptr;
- res = LLVMBuildBitCast(gallivm->builder, res, v2i128, "");
- res = LLVMBuildExtractElement(gallivm->builder, res, bld_base->uint_bld.one, "");
- res = LLVMBuildBitCast(gallivm->builder, res, ctx->v16i8, "");
-
emit_data->dst_type = ctx->v4f32;
- emit_data->args[0] = res;
+ emit_data->args[0] = LLVMBuildBitCast(gallivm->builder, res_ptr,
+ ctx->v16i8, "");
emit_data->args[1] = bld_base->uint_bld.zero;
emit_data->args[2] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
emit_data->arg_count = 3;
emit_data->output[emit_data->chan] =
lp_build_intrinsic(builder, intr_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
}
static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
base->gallivm->builder,
"llvm.SI.vs.load.input", emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
return;
}
emit_data->output[emit_data->chan] = lp_build_intrinsic(
base->gallivm->builder, intr_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute);
+ LP_FUNC_ATTR_READNONE);
}
static void si_llvm_emit_txqs(
args[1] = val;
tl = lp_build_intrinsic(gallivm->builder,
"llvm.amdgcn.ds.bpermute", ctx->i32,
- args, 2, LLVMReadNoneAttribute);
+ args, 2, LP_FUNC_ATTR_READNONE);
args[0] = LLVMBuildMul(gallivm->builder, trbl_tid,
lp_build_const_int32(gallivm, 4), "");
trbl = lp_build_intrinsic(gallivm->builder,
"llvm.amdgcn.ds.bpermute", ctx->i32,
- args, 2, LLVMReadNoneAttribute);
+ args, 2, LP_FUNC_ATTR_READNONE);
} else {
LLVMValueRef store_ptr, load_ptr0, load_ptr1;
struct si_shader_context *ctx = si_shader_context(bld_base);
struct si_shader *shader = ctx->shader;
struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct lp_build_context *uint = &bld_base->uint_bld;
LLVMValueRef interp_param;
const struct tgsi_full_instruction *inst = emit_data->inst;
- const char *intr_name;
int input_index = inst->Src[0].Register.Index;
int chan;
int i;
if (interp_param_idx == -1)
return;
else if (interp_param_idx)
- interp_param = get_interp_param(ctx, interp_param_idx);
+ interp_param = LLVMGetParam(ctx->main_fn, interp_param_idx);
else
interp_param = NULL;
temp2 = LLVMBuildFMul(gallivm->builder, ddy_el, emit_data->args[1], "");
- temp2 = LLVMBuildFAdd(gallivm->builder, temp2, temp1, "");
-
- ij_out[i] = LLVMBuildBitCast(gallivm->builder,
- temp2, ctx->i32, "");
+ ij_out[i] = LLVMBuildFAdd(gallivm->builder, temp2, temp1, "");
}
interp_param = lp_build_gather_values(bld_base->base.gallivm, ij_out, 2);
}
- intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
for (chan = 0; chan < 4; chan++) {
- LLVMValueRef args[4];
LLVMValueRef llvm_chan;
unsigned schan;
schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan);
llvm_chan = lp_build_const_int32(gallivm, schan);
- args[0] = llvm_chan;
- args[1] = attr_number;
- args[2] = params;
- args[3] = interp_param;
-
- emit_data->output[chan] =
- lp_build_intrinsic(gallivm->builder, intr_name,
- ctx->f32, args, args[3] ? 4 : 3,
- LLVMReadNoneAttribute);
+ if (interp_param) {
+ interp_param = LLVMBuildBitCast(gallivm->builder,
+ interp_param, LLVMVectorType(ctx->f32, 2), "");
+ LLVMValueRef i = LLVMBuildExtractElement(
+ gallivm->builder, interp_param, uint->zero, "");
+ LLVMValueRef j = LLVMBuildExtractElement(
+ gallivm->builder, interp_param, uint->one, "");
+ emit_data->output[chan] = build_fs_interp(bld_base,
+ llvm_chan, attr_number, params,
+ i, j);
+ } else {
+ emit_data->output[chan] = build_fs_interp_mov(bld_base,
+ lp_build_const_int32(gallivm, 2), /* P0 */
+ llvm_chan, attr_number, params);
+ }
}
}
struct si_shader *shader = ctx->shader;
struct tgsi_shader_info *info = &shader->selector->info;
struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct lp_build_if_state if_state;
LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
SI_PARAM_GS2VS_OFFSET);
LLVMValueRef gs_next_vertex;
"");
/* If this thread has already emitted the declared maximum number of
- * vertices, kill it: excessive vertex emissions are not supposed to
- * have any effect, and GS threads have no externally observable
- * effects other than emitting vertices.
+ * vertices, skip the write: excessive vertex emissions are not
+ * supposed to have any effect.
+ *
+ * If the shader has no writes to memory, kill it instead. This skips
+ * further memory loads and may allow LLVM to skip to the end
+ * altogether.
*/
- can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULE, gs_next_vertex,
+ can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULT, gs_next_vertex,
lp_build_const_int32(gallivm,
shader->selector->gs_max_out_vertices), "");
- kill = lp_build_select(&bld_base->base, can_emit,
- lp_build_const_float(gallivm, 1.0f),
- lp_build_const_float(gallivm, -1.0f));
- lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
- ctx->voidt, &kill, 1, 0);
+ bool use_kill = !info->writes_memory;
+ if (use_kill) {
+ kill = lp_build_select(&bld_base->base, can_emit,
+ lp_build_const_float(gallivm, 1.0f),
+ lp_build_const_float(gallivm, -1.0f));
+
+ lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
+ ctx->voidt, &kill, 1, 0);
+ } else {
+ lp_build_if(&if_state, gallivm, can_emit);
+ }
for (i = 0; i < info->num_outputs; i++) {
LLVMValueRef *out_ptr =
ctx->soa.outputs[i];
for (chan = 0; chan < 4; chan++) {
+ if (((info->output_streams[i] >> (2 * chan)) & 3) != stream)
+ continue;
+
LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
LLVMValueRef voffset =
lp_build_const_int32(gallivm, (i * 4 + chan) *
1, 0, 1, 1, 0);
}
}
+
gs_next_vertex = lp_build_add(uint, gs_next_vertex,
lp_build_const_int32(gallivm, 1));
args[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID);
lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
ctx->voidt, args, 2, 0);
+
+ if (!use_kill)
+ lp_build_endif(&if_state);
}
/* Cut one primitive from the geometry shader */
* always fits into a single wave.
*/
if (ctx->type == PIPE_SHADER_TESS_CTRL) {
- emit_optimization_barrier(ctx);
+ emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
return;
}
* SGPR spilling significantly.
*/
if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
- LLVMAddAttribute(P, LLVMByValAttribute);
+ lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_BYVAL);
lp_add_attr_dereferenceable(P, UINT64_MAX);
} else
- LLVMAddAttribute(P, LLVMInRegAttribute);
+ lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_INREG);
}
if (ctx->screen->b.debug_flags & DBG_UNSAFE_MATH) {
case LLVMVectorTypeKind:
return LLVMGetVectorSize(type) *
llvm_get_type_size(LLVMGetElementType(type));
+ case LLVMArrayTypeKind:
+ return LLVMGetArrayLength(type) *
+ llvm_get_type_size(LLVMGetElementType(type));
default:
assert(0);
return 0;
"tess_lds");
}
+static unsigned si_get_max_workgroup_size(struct si_shader *shader)
+{
+ const unsigned *properties = shader->selector->info.properties;
+ unsigned max_work_group_size =
+ properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
+ properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
+ properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
+
+ if (!max_work_group_size) {
+ /* This is a variable group size compute shader,
+ * compile it for the maximum possible group size.
+ */
+ max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
+ }
+ return max_work_group_size;
+}
+
static void create_function(struct si_shader_context *ctx)
{
struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
params[SI_PARAM_DRAWID] = ctx->i32;
num_params = SI_PARAM_DRAWID+1;
- if (shader->key.vs.as_es) {
+ if (shader->key.as_es) {
params[ctx->param_es2gs_offset = num_params++] = ctx->i32;
- } else if (shader->key.vs.as_ls) {
+ } else if (shader->key.as_ls) {
params[SI_PARAM_LS_OUT_LAYOUT] = ctx->i32;
num_params = SI_PARAM_LS_OUT_LAYOUT+1;
} else {
- if (ctx->is_gs_copy_shader) {
+ if (shader->is_gs_copy_shader) {
num_params = SI_PARAM_RW_BUFFERS+1;
} else {
params[SI_PARAM_VS_STATE_BITS] = ctx->i32;
params[ctx->param_vs_prim_id = num_params++] = ctx->i32;
params[ctx->param_instance_id = num_params++] = ctx->i32;
- if (!ctx->no_prolog &&
- !ctx->is_gs_copy_shader) {
+ if (!shader->is_gs_copy_shader) {
/* Vertex load indices. */
ctx->param_vertex_index0 = num_params;
params[num_params++] = ctx->i32;
num_prolog_vgprs += shader->selector->info.num_inputs;
- }
- if (!ctx->no_epilog &&
- !ctx->is_gs_copy_shader) {
/* PrimitiveID output. */
- if (!shader->key.vs.as_es && !shader->key.vs.as_ls)
+ if (!shader->key.as_es && !shader->key.as_ls)
for (i = 0; i <= VS_EPILOG_PRIMID_LOC; i++)
returns[num_returns++] = ctx->f32;
}
params[SI_PARAM_REL_IDS] = ctx->i32;
num_params = SI_PARAM_REL_IDS+1;
- if (!ctx->no_epilog) {
- /* SI_PARAM_TCS_OC_LDS and PARAM_TESS_FACTOR_OFFSET are
- * placed after the user SGPRs.
- */
- for (i = 0; i < SI_TCS_NUM_USER_SGPR + 2; i++)
- returns[num_returns++] = ctx->i32; /* SGPRs */
+ /* SI_PARAM_TCS_OC_LDS and PARAM_TESS_FACTOR_OFFSET are
+ * placed after the user SGPRs.
+ */
+ for (i = 0; i < SI_TCS_NUM_USER_SGPR + 2; i++)
+ returns[num_returns++] = ctx->i32; /* SGPRs */
- for (i = 0; i < 3; i++)
- returns[num_returns++] = ctx->f32; /* VGPRs */
- }
+ for (i = 0; i < 3; i++)
+ returns[num_returns++] = ctx->f32; /* VGPRs */
break;
case PIPE_SHADER_TESS_EVAL:
params[SI_PARAM_TCS_OFFCHIP_LAYOUT] = ctx->i32;
num_params = SI_PARAM_TCS_OFFCHIP_LAYOUT+1;
- if (shader->key.tes.as_es) {
+ if (shader->key.as_es) {
params[ctx->param_oc_lds = num_params++] = ctx->i32;
params[ctx->param_tess_offchip = num_params++] = ctx->i32;
params[ctx->param_es2gs_offset = num_params++] = ctx->i32;
params[ctx->param_tes_patch_id = num_params++] = ctx->i32;
/* PrimitiveID output. */
- if (!ctx->no_epilog && !shader->key.tes.as_es)
+ if (!shader->key.as_es)
for (i = 0; i <= VS_EPILOG_PRIMID_LOC; i++)
returns[num_returns++] = ctx->f32;
break;
params[SI_PARAM_POS_FIXED_PT] = ctx->i32;
num_params = SI_PARAM_POS_FIXED_PT+1;
- if (!ctx->no_prolog) {
- /* Color inputs from the prolog. */
- if (shader->selector->info.colors_read) {
- unsigned num_color_elements =
- util_bitcount(shader->selector->info.colors_read);
+ /* Color inputs from the prolog. */
+ if (shader->selector->info.colors_read) {
+ unsigned num_color_elements =
+ util_bitcount(shader->selector->info.colors_read);
- assert(num_params + num_color_elements <= ARRAY_SIZE(params));
- for (i = 0; i < num_color_elements; i++)
- params[num_params++] = ctx->f32;
+ assert(num_params + num_color_elements <= ARRAY_SIZE(params));
+ for (i = 0; i < num_color_elements; i++)
+ params[num_params++] = ctx->f32;
- num_prolog_vgprs += num_color_elements;
- }
+ num_prolog_vgprs += num_color_elements;
}
- if (!ctx->no_epilog) {
- /* Outputs for the epilog. */
- num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
- num_returns =
- num_return_sgprs +
- util_bitcount(shader->selector->info.colors_written) * 4 +
- shader->selector->info.writes_z +
- shader->selector->info.writes_stencil +
- shader->selector->info.writes_samplemask +
- 1 /* SampleMaskIn */;
-
- num_returns = MAX2(num_returns,
- num_return_sgprs +
- PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
-
- for (i = 0; i < num_return_sgprs; i++)
- returns[i] = ctx->i32;
- for (; i < num_returns; i++)
- returns[i] = ctx->f32;
- }
+ /* Outputs for the epilog. */
+ num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
+ num_returns =
+ num_return_sgprs +
+ util_bitcount(shader->selector->info.colors_written) * 4 +
+ shader->selector->info.writes_z +
+ shader->selector->info.writes_stencil +
+ shader->selector->info.writes_samplemask +
+ 1 /* SampleMaskIn */;
+
+ num_returns = MAX2(num_returns,
+ num_return_sgprs +
+ PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
+
+ for (i = 0; i < num_return_sgprs; i++)
+ returns[i] = ctx->i32;
+ for (; i < num_returns; i++)
+ returns[i] = ctx->f32;
break;
case PIPE_SHADER_COMPUTE:
S_0286D0_FRONT_FACE_ENA(1) |
S_0286D0_POS_FIXED_PT_ENA(1));
} else if (ctx->type == PIPE_SHADER_COMPUTE) {
- const unsigned *properties = shader->selector->info.properties;
- unsigned max_work_group_size =
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
-
- if (!max_work_group_size) {
- /* This is a variable group size compute shader,
- * compile it for the maximum possible group size.
- */
- max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
- }
-
si_llvm_add_attribute(ctx->main_fn,
"amdgpu-max-work-group-size",
- max_work_group_size);
+ si_get_max_workgroup_size(shader));
}
shader->info.num_input_sgprs = 0;
"ddxy_lds",
LOCAL_ADDR_SPACE);
- if ((ctx->type == PIPE_SHADER_VERTEX && shader->key.vs.as_ls) ||
+ if ((ctx->type == PIPE_SHADER_VERTEX && shader->key.as_ls) ||
ctx->type == PIPE_SHADER_TESS_CTRL ||
ctx->type == PIPE_SHADER_TESS_EVAL)
declare_tess_lds(ctx);
SI_PARAM_RW_BUFFERS);
if ((ctx->type == PIPE_SHADER_VERTEX &&
- ctx->shader->key.vs.as_es) ||
+ ctx->shader->key.as_es) ||
(ctx->type == PIPE_SHADER_TESS_EVAL &&
- ctx->shader->key.tes.as_es) ||
+ ctx->shader->key.as_es) ||
ctx->type == PIPE_SHADER_GEOMETRY) {
unsigned ring =
ctx->type == PIPE_SHADER_GEOMETRY ? SI_GS_RING_ESGS
build_indexed_load_const(ctx, buf_ptr, offset);
}
- if (ctx->is_gs_copy_shader) {
+ if (ctx->shader->is_gs_copy_shader) {
LLVMValueRef offset = lp_build_const_int32(gallivm, SI_VS_RING_GSVS);
ctx->gsvs_ring[0] =
}
static void si_shader_dump_stats(struct si_screen *sscreen,
- struct si_shader_config *conf,
- unsigned num_inputs,
- unsigned code_size,
+ struct si_shader *shader,
struct pipe_debug_callback *debug,
unsigned processor,
FILE *file)
{
+ struct si_shader_config *conf = &shader->config;
+ unsigned num_inputs = shader->selector ? shader->selector->info.num_inputs : 0;
+ unsigned code_size = si_get_shader_binary_size(shader);
unsigned lds_increment = sscreen->b.chip_class >= CIK ? 512 : 256;
unsigned lds_per_wave = 0;
unsigned max_simd_waves = 10;
/* Compute LDS usage for PS. */
- if (processor == PIPE_SHADER_FRAGMENT) {
+ switch (processor) {
+ case PIPE_SHADER_FRAGMENT:
/* The minimum usage per wave is (num_inputs * 48). The maximum
* usage is (num_inputs * 48 * 16).
* We can get anything in between and it varies between waves.
*/
lds_per_wave = conf->lds_size * lds_increment +
align(num_inputs * 48, lds_increment);
+ break;
+ case PIPE_SHADER_COMPUTE:
+ if (shader->selector) {
+ unsigned max_workgroup_size =
+ si_get_max_workgroup_size(shader);
+ lds_per_wave = (conf->lds_size * lds_increment) /
+ DIV_ROUND_UP(max_workgroup_size, 64);
+ }
+ break;
}
/* Compute the per-SIMD wave counts. */
if (conf->num_vgprs)
max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);
- /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
- * that PS can use.
- */
+ /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
+ * 16KB makes some SIMDs unoccupied). */
if (lds_per_wave)
max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
"VGPRS: %d\n"
"Spilled SGPRs: %d\n"
"Spilled VGPRs: %d\n"
+ "Private memory VGPRs: %d\n"
"Code Size: %d bytes\n"
"LDS: %d blocks\n"
"Scratch: %d bytes per wave\n"
"Max Waves: %d\n"
"********************\n\n\n",
conf->num_sgprs, conf->num_vgprs,
- conf->spilled_sgprs, conf->spilled_vgprs, code_size,
+ conf->spilled_sgprs, conf->spilled_vgprs,
+ conf->private_mem_vgprs, code_size,
conf->lds_size, conf->scratch_bytes_per_wave,
max_simd_waves);
}
pipe_debug_message(debug, SHADER_INFO,
"Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
"LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
- "Spilled VGPRs: %d",
+ "Spilled VGPRs: %d PrivMem VGPRs: %d",
conf->num_sgprs, conf->num_vgprs, code_size,
conf->lds_size, conf->scratch_bytes_per_wave,
max_simd_waves, conf->spilled_sgprs,
- conf->spilled_vgprs);
+ conf->spilled_vgprs, conf->private_mem_vgprs);
}
static const char *si_get_shader_name(struct si_shader *shader,
{
switch (processor) {
case PIPE_SHADER_VERTEX:
- if (shader->key.vs.as_es)
+ if (shader->key.as_es)
return "Vertex Shader as ES";
- else if (shader->key.vs.as_ls)
+ else if (shader->key.as_ls)
return "Vertex Shader as LS";
else
return "Vertex Shader as VS";
case PIPE_SHADER_TESS_CTRL:
return "Tessellation Control Shader";
case PIPE_SHADER_TESS_EVAL:
- if (shader->key.tes.as_es)
+ if (shader->key.as_es)
return "Tessellation Evaluation Shader as ES";
else
return "Tessellation Evaluation Shader as VS";
case PIPE_SHADER_GEOMETRY:
- if (shader->gs_copy_shader == NULL)
+ if (shader->is_gs_copy_shader)
return "GS Copy Shader as VS";
else
return "Geometry Shader";
fprintf(file, "\n");
}
- si_shader_dump_stats(sscreen, &shader->config,
- shader->selector ? shader->selector->info.num_inputs : 0,
- si_get_shader_binary_size(shader), debug, processor,
- file);
+ si_shader_dump_stats(sscreen, shader, debug, processor, file);
}
int si_compile_llvm(struct si_screen *sscreen,
}
/* Generate code for the hardware VS shader stage to go with a geometry shader */
-static int si_generate_gs_copy_shader(struct si_screen *sscreen,
- struct si_shader_context *ctx,
- struct si_shader *gs,
- struct pipe_debug_callback *debug)
+struct si_shader *
+si_generate_gs_copy_shader(struct si_screen *sscreen,
+ LLVMTargetMachineRef tm,
+ struct si_shader_selector *gs_selector,
+ struct pipe_debug_callback *debug)
{
- struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct si_shader_context ctx;
+ struct si_shader *shader;
+ struct gallivm_state *gallivm = &ctx.gallivm;
+ LLVMBuilderRef builder;
+ struct lp_build_tgsi_context *bld_base = &ctx.soa.bld_base;
struct lp_build_context *uint = &bld_base->uint_bld;
struct si_shader_output_values *outputs;
- struct tgsi_shader_info *gsinfo = &gs->selector->info;
+ struct tgsi_shader_info *gsinfo = &gs_selector->info;
LLVMValueRef args[9];
int i, r;
outputs = MALLOC(gsinfo->num_outputs * sizeof(outputs[0]));
- si_init_shader_ctx(ctx, sscreen, ctx->shader, ctx->tm);
- ctx->type = PIPE_SHADER_VERTEX;
- ctx->is_gs_copy_shader = true;
+ if (!outputs)
+ return NULL;
- create_meta_data(ctx);
- create_function(ctx);
- preload_ring_buffers(ctx);
+ shader = CALLOC_STRUCT(si_shader);
+ if (!shader) {
+ FREE(outputs);
+ return NULL;
+ }
+
+
+ shader->selector = gs_selector;
+ shader->is_gs_copy_shader = true;
+
+ si_init_shader_ctx(&ctx, sscreen, shader, tm);
+ ctx.type = PIPE_SHADER_VERTEX;
- args[0] = ctx->gsvs_ring[0];
+ builder = gallivm->builder;
+
+ create_meta_data(&ctx);
+ create_function(&ctx);
+ preload_ring_buffers(&ctx);
+
+ args[0] = ctx.gsvs_ring[0];
args[1] = lp_build_mul_imm(uint,
- LLVMGetParam(ctx->main_fn,
- ctx->param_vertex_id),
+ LLVMGetParam(ctx.main_fn,
+ ctx.param_vertex_id),
4);
args[3] = uint->zero;
args[4] = uint->one; /* OFFEN */
args[7] = uint->one; /* SLC */
args[8] = uint->zero; /* TFE */
- /* Fetch vertex data from GSVS ring */
+ /* Fetch the vertex stream ID.*/
+ LLVMValueRef stream_id;
+
+ if (gs_selector->so.num_outputs)
+ stream_id = unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
+ else
+ stream_id = uint->zero;
+
+ /* Fill in output information. */
for (i = 0; i < gsinfo->num_outputs; ++i) {
- unsigned chan;
+ outputs[i].semantic_name = gsinfo->output_semantic_name[i];
+ outputs[i].semantic_index = gsinfo->output_semantic_index[i];
+
+ for (int chan = 0; chan < 4; chan++) {
+ outputs[i].vertex_stream[chan] =
+ (gsinfo->output_streams[i] >> (2 * chan)) & 3;
+ }
+ }
- outputs[i].name = gsinfo->output_semantic_name[i];
- outputs[i].sid = gsinfo->output_semantic_index[i];
+ LLVMBasicBlockRef end_bb;
+ LLVMValueRef switch_inst;
- for (chan = 0; chan < 4; chan++) {
- args[2] = lp_build_const_int32(gallivm,
- (i * 4 + chan) *
- gs->selector->gs_max_out_vertices * 16 * 4);
+ end_bb = LLVMAppendBasicBlockInContext(gallivm->context, ctx.main_fn, "end");
+ switch_inst = LLVMBuildSwitch(builder, stream_id, end_bb, 4);
+
+ for (int stream = 0; stream < 4; stream++) {
+ LLVMBasicBlockRef bb;
+
+ if (!gsinfo->num_stream_output_components[stream])
+ continue;
+
+ if (stream > 0 && !gs_selector->so.num_outputs)
+ continue;
- outputs[i].values[chan] =
- LLVMBuildBitCast(gallivm->builder,
+ bb = LLVMInsertBasicBlockInContext(gallivm->context, end_bb, "out");
+ LLVMAddCase(switch_inst, lp_build_const_int32(gallivm, stream), bb);
+ LLVMPositionBuilderAtEnd(builder, bb);
+
+ /* Fetch vertex data from GSVS ring */
+ for (i = 0; i < gsinfo->num_outputs; ++i) {
+ for (unsigned chan = 0; chan < 4; chan++) {
+ if (outputs[i].vertex_stream[chan] != stream) {
+ outputs[i].values[chan] = ctx.soa.bld_base.base.undef;
+ continue;
+ }
+
+ args[2] = lp_build_const_int32(
+ gallivm,
+ (i * 4 + chan) * gs_selector->gs_max_out_vertices * 16 * 4);
+
+ outputs[i].values[chan] =
+ LLVMBuildBitCast(gallivm->builder,
lp_build_intrinsic(gallivm->builder,
"llvm.SI.buffer.load.dword.i32.i32",
- ctx->i32, args, 9,
- LLVMReadOnlyAttribute),
- ctx->f32, "");
+ ctx.i32, args, 9,
+ LP_FUNC_ATTR_READONLY),
+ ctx.f32, "");
+ }
}
+
+ /* Streamout and exports. */
+ if (gs_selector->so.num_outputs) {
+ si_llvm_emit_streamout(&ctx, outputs,
+ gsinfo->num_outputs,
+ stream);
+ }
+
+ if (stream == 0)
+ si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
+
+ LLVMBuildBr(builder, end_bb);
}
- si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
+ LLVMPositionBuilderAtEnd(builder, end_bb);
LLVMBuildRetVoid(gallivm->builder);
r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
LLVMDumpModule(bld_base->base.gallivm->module);
- si_llvm_finalize_module(ctx,
+ si_llvm_finalize_module(&ctx,
r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_GEOMETRY));
- r = si_compile_llvm(sscreen, &ctx->shader->binary,
- &ctx->shader->config, ctx->tm,
+ r = si_compile_llvm(sscreen, &ctx.shader->binary,
+ &ctx.shader->config, ctx.tm,
bld_base->base.gallivm->module,
debug, PIPE_SHADER_GEOMETRY,
"GS Copy Shader");
if (!r) {
if (r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
fprintf(stderr, "GS Copy Shader:\n");
- si_shader_dump(sscreen, ctx->shader, debug,
+ si_shader_dump(sscreen, ctx.shader, debug,
PIPE_SHADER_GEOMETRY, stderr);
- r = si_shader_binary_upload(sscreen, ctx->shader);
+ r = si_shader_binary_upload(sscreen, ctx.shader);
}
- si_llvm_dispose(ctx);
+ si_llvm_dispose(&ctx);
FREE(outputs);
- return r;
+
+ if (r != 0) {
+ FREE(shader);
+ shader = NULL;
+ }
+ return shader;
}
-static void si_dump_shader_key(unsigned shader, union si_shader_key *key,
+static void si_dump_shader_key(unsigned shader, struct si_shader_key *key,
FILE *f)
{
int i;
switch (shader) {
case PIPE_SHADER_VERTEX:
- fprintf(f, " instance_divisors = {");
- for (i = 0; i < ARRAY_SIZE(key->vs.prolog.instance_divisors); i++)
+ fprintf(f, " part.vs.prolog.instance_divisors = {");
+ for (i = 0; i < ARRAY_SIZE(key->part.vs.prolog.instance_divisors); i++)
fprintf(f, !i ? "%u" : ", %u",
- key->vs.prolog.instance_divisors[i]);
+ key->part.vs.prolog.instance_divisors[i]);
fprintf(f, "}\n");
- fprintf(f, " as_es = %u\n", key->vs.as_es);
- fprintf(f, " as_ls = %u\n", key->vs.as_ls);
- fprintf(f, " export_prim_id = %u\n", key->vs.epilog.export_prim_id);
+ fprintf(f, " part.vs.epilog.export_prim_id = %u\n", key->part.vs.epilog.export_prim_id);
+ fprintf(f, " as_es = %u\n", key->as_es);
+ fprintf(f, " as_ls = %u\n", key->as_ls);
+ fprintf(f, " mono.vs.fix_fetch = 0x%x\n", key->mono.vs.fix_fetch);
break;
case PIPE_SHADER_TESS_CTRL:
- fprintf(f, " prim_mode = %u\n", key->tcs.epilog.prim_mode);
+ fprintf(f, " part.tcs.epilog.prim_mode = %u\n", key->part.tcs.epilog.prim_mode);
+ fprintf(f, " mono.tcs.inputs_to_copy = 0x%"PRIx64"\n", key->mono.tcs.inputs_to_copy);
break;
case PIPE_SHADER_TESS_EVAL:
- fprintf(f, " as_es = %u\n", key->tes.as_es);
- fprintf(f, " export_prim_id = %u\n", key->tes.epilog.export_prim_id);
+ fprintf(f, " part.tes.epilog.export_prim_id = %u\n", key->part.tes.epilog.export_prim_id);
+ fprintf(f, " as_es = %u\n", key->as_es);
break;
case PIPE_SHADER_GEOMETRY:
+ fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
+ break;
+
case PIPE_SHADER_COMPUTE:
break;
case PIPE_SHADER_FRAGMENT:
- fprintf(f, " prolog.color_two_side = %u\n", key->ps.prolog.color_two_side);
- fprintf(f, " prolog.flatshade_colors = %u\n", key->ps.prolog.flatshade_colors);
- fprintf(f, " prolog.poly_stipple = %u\n", key->ps.prolog.poly_stipple);
- fprintf(f, " prolog.force_persp_sample_interp = %u\n", key->ps.prolog.force_persp_sample_interp);
- fprintf(f, " prolog.force_linear_sample_interp = %u\n", key->ps.prolog.force_linear_sample_interp);
- fprintf(f, " prolog.force_persp_center_interp = %u\n", key->ps.prolog.force_persp_center_interp);
- fprintf(f, " prolog.force_linear_center_interp = %u\n", key->ps.prolog.force_linear_center_interp);
- fprintf(f, " prolog.bc_optimize_for_persp = %u\n", key->ps.prolog.bc_optimize_for_persp);
- fprintf(f, " prolog.bc_optimize_for_linear = %u\n", key->ps.prolog.bc_optimize_for_linear);
- fprintf(f, " epilog.spi_shader_col_format = 0x%x\n", key->ps.epilog.spi_shader_col_format);
- fprintf(f, " epilog.color_is_int8 = 0x%X\n", key->ps.epilog.color_is_int8);
- fprintf(f, " epilog.last_cbuf = %u\n", key->ps.epilog.last_cbuf);
- fprintf(f, " epilog.alpha_func = %u\n", key->ps.epilog.alpha_func);
- fprintf(f, " epilog.alpha_to_one = %u\n", key->ps.epilog.alpha_to_one);
- fprintf(f, " epilog.poly_line_smoothing = %u\n", key->ps.epilog.poly_line_smoothing);
- fprintf(f, " epilog.clamp_color = %u\n", key->ps.epilog.clamp_color);
+ fprintf(f, " part.ps.prolog.color_two_side = %u\n", key->part.ps.prolog.color_two_side);
+ fprintf(f, " part.ps.prolog.flatshade_colors = %u\n", key->part.ps.prolog.flatshade_colors);
+ fprintf(f, " part.ps.prolog.poly_stipple = %u\n", key->part.ps.prolog.poly_stipple);
+ fprintf(f, " part.ps.prolog.force_persp_sample_interp = %u\n", key->part.ps.prolog.force_persp_sample_interp);
+ fprintf(f, " part.ps.prolog.force_linear_sample_interp = %u\n", key->part.ps.prolog.force_linear_sample_interp);
+ fprintf(f, " part.ps.prolog.force_persp_center_interp = %u\n", key->part.ps.prolog.force_persp_center_interp);
+ fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
+ fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
+ fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
+ fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
+ fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
+ fprintf(f, " part.ps.epilog.last_cbuf = %u\n", key->part.ps.epilog.last_cbuf);
+ fprintf(f, " part.ps.epilog.alpha_func = %u\n", key->part.ps.epilog.alpha_func);
+ fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
+ fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
+ fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
break;
default:
assert(0);
}
+
+ if ((shader == PIPE_SHADER_GEOMETRY ||
+ shader == PIPE_SHADER_TESS_EVAL ||
+ shader == PIPE_SHADER_VERTEX) &&
+ !key->as_es && !key->as_ls) {
+ fprintf(f, " opt.hw_vs.kill_outputs = 0x%"PRIx64"\n", key->opt.hw_vs.kill_outputs);
+ fprintf(f, " opt.hw_vs.kill_outputs2 = 0x%x\n", key->opt.hw_vs.kill_outputs2);
+ fprintf(f, " opt.hw_vs.clip_disable = %u\n", key->opt.hw_vs.clip_disable);
+ }
}
static void si_init_shader_ctx(struct si_shader_context *ctx,
struct lp_build_tgsi_context *bld_base;
struct lp_build_tgsi_action tmpl = {};
- memset(ctx, 0, sizeof(*ctx));
- si_llvm_context_init(
- ctx, "amdgcn--",
+ si_llvm_context_init(ctx, sscreen, shader, tm,
(shader && shader->selector) ? &shader->selector->info : NULL,
(shader && shader->selector) ? shader->selector->tokens : NULL);
- si_shader_context_init_alu(&ctx->soa.bld_base);
- ctx->tm = tm;
- ctx->screen = sscreen;
- if (shader && shader->selector)
- ctx->type = shader->selector->info.processor;
- else
- ctx->type = -1;
- ctx->shader = shader;
-
- ctx->voidt = LLVMVoidTypeInContext(ctx->gallivm.context);
- ctx->i1 = LLVMInt1TypeInContext(ctx->gallivm.context);
- ctx->i8 = LLVMInt8TypeInContext(ctx->gallivm.context);
- ctx->i32 = LLVMInt32TypeInContext(ctx->gallivm.context);
- ctx->i64 = LLVMInt64TypeInContext(ctx->gallivm.context);
- ctx->i128 = LLVMIntTypeInContext(ctx->gallivm.context, 128);
- ctx->f32 = LLVMFloatTypeInContext(ctx->gallivm.context);
- ctx->v16i8 = LLVMVectorType(ctx->i8, 16);
- ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
- ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
- ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
- ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
bld_base = &ctx->soa.bld_base;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
exports.num = 0;
- if ((ctx->type == PIPE_SHADER_VERTEX &&
- (shader->key.vs.as_es || shader->key.vs.as_ls)) ||
- (ctx->type == PIPE_SHADER_TESS_EVAL && shader->key.tes.as_es))
+ if (ctx->type == PIPE_SHADER_FRAGMENT ||
+ ctx->type == PIPE_SHADER_COMPUTE ||
+ shader->key.as_es ||
+ shader->key.as_ls)
return;
/* Process all LLVM instructions. */
}
}
+static void si_count_scratch_private_memory(struct si_shader_context *ctx)
+{
+ ctx->shader->config.private_mem_vgprs = 0;
+
+ /* Process all LLVM instructions. */
+ LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(ctx->main_fn);
+ while (bb) {
+ LLVMValueRef next = LLVMGetFirstInstruction(bb);
+
+ while (next) {
+ LLVMValueRef inst = next;
+ next = LLVMGetNextInstruction(next);
+
+ if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
+ continue;
+
+ LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
+ /* No idea why LLVM aligns allocas to 4 elements. */
+ unsigned alignment = LLVMGetAlignment(inst);
+ unsigned dw_size = align(llvm_get_type_size(type) / 4, alignment);
+ ctx->shader->config.private_mem_vgprs += dw_size;
+ }
+ bb = LLVMGetNextBasicBlock(bb);
+ }
+}
+
static bool si_compile_tgsi_main(struct si_shader_context *ctx,
struct si_shader *shader)
{
switch (ctx->type) {
case PIPE_SHADER_VERTEX:
ctx->load_input = declare_input_vs;
- if (shader->key.vs.as_ls)
+ if (shader->key.as_ls)
bld_base->emit_epilogue = si_llvm_emit_ls_epilogue;
- else if (shader->key.vs.as_es)
+ else if (shader->key.as_es)
bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
else
bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
break;
case PIPE_SHADER_TESS_EVAL:
bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tes;
- if (shader->key.tes.as_es)
+ if (shader->key.as_es)
bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
else
bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
break;
case PIPE_SHADER_FRAGMENT:
ctx->load_input = declare_input_fs;
- if (ctx->no_epilog)
- bld_base->emit_epilogue = si_llvm_emit_fs_epilogue;
- else
- bld_base->emit_epilogue = si_llvm_return_fs_outputs;
+ bld_base->emit_epilogue = si_llvm_return_fs_outputs;
break;
case PIPE_SHADER_COMPUTE:
ctx->declare_memory_region = declare_compute_memory;
create_function(ctx);
preload_ring_buffers(ctx);
- if (ctx->no_prolog && sel->type == PIPE_SHADER_FRAGMENT &&
- shader->key.ps.prolog.poly_stipple) {
- LLVMValueRef list = LLVMGetParam(ctx->main_fn,
- SI_PARAM_RW_BUFFERS);
- si_llvm_emit_polygon_stipple(ctx, list,
- SI_PARAM_POS_FIXED_PT);
- }
-
if (ctx->type == PIPE_SHADER_GEOMETRY) {
int i;
for (i = 0; i < 4; i++) {
struct tgsi_shader_info *info = &shader->selector->info;
memset(key, 0, sizeof(*key));
- key->vs_prolog.states = shader->key.vs.prolog;
+ key->vs_prolog.states = shader->key.part.vs.prolog;
key->vs_prolog.num_input_sgprs = shader->info.num_input_sgprs;
key->vs_prolog.last_input = MAX2(1, info->num_inputs) - 1;
key->vs_epilog.states = *states;
/* Set up the PrimitiveID output. */
- if (shader->key.vs.epilog.export_prim_id) {
+ if (shader->key.part.vs.epilog.export_prim_id) {
unsigned index = shader->selector->info.num_outputs;
unsigned offset = shader->info.nr_param_exports++;
struct tgsi_shader_info *info = &shader->selector->info;
memset(key, 0, sizeof(*key));
- key->ps_prolog.states = shader->key.ps.prolog;
+ key->ps_prolog.states = shader->key.part.ps.prolog;
key->ps_prolog.colors_read = info->colors_read;
key->ps_prolog.num_input_sgprs = shader->info.num_input_sgprs;
key->ps_prolog.num_input_vgprs = shader->info.num_input_vgprs;
if (info->colors_read) {
unsigned *color = shader->selector->color_attr_index;
- if (shader->key.ps.prolog.color_two_side) {
+ if (shader->key.part.ps.prolog.color_two_side) {
/* BCOLORs are stored after the last input. */
key->ps_prolog.num_interp_inputs = info->num_inputs;
key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
key->ps_prolog.color_attr_index[i] = color[i];
- if (shader->key.ps.prolog.flatshade_colors &&
+ if (shader->key.part.ps.prolog.flatshade_colors &&
interp == TGSI_INTERPOLATE_COLOR)
interp = TGSI_INTERPOLATE_CONSTANT;
case TGSI_INTERPOLATE_PERSPECTIVE:
case TGSI_INTERPOLATE_COLOR:
/* Force the interpolation location for colors here. */
- if (shader->key.ps.prolog.force_persp_sample_interp)
+ if (shader->key.part.ps.prolog.force_persp_sample_interp)
location = TGSI_INTERPOLATE_LOC_SAMPLE;
- if (shader->key.ps.prolog.force_persp_center_interp)
+ if (shader->key.part.ps.prolog.force_persp_center_interp)
location = TGSI_INTERPOLATE_LOC_CENTER;
switch (location) {
break;
case TGSI_INTERPOLATE_LINEAR:
/* Force the interpolation location for colors here. */
- if (shader->key.ps.prolog.force_linear_sample_interp)
+ if (shader->key.part.ps.prolog.force_linear_sample_interp)
location = TGSI_INTERPOLATE_LOC_SAMPLE;
- if (shader->key.ps.prolog.force_linear_center_interp)
+ if (shader->key.part.ps.prolog.force_linear_center_interp)
location = TGSI_INTERPOLATE_LOC_CENTER;
/* The VGPR assignment for non-monolithic shaders
key->ps_epilog.writes_z = info->writes_z;
key->ps_epilog.writes_stencil = info->writes_stencil;
key->ps_epilog.writes_samplemask = info->writes_samplemask;
- key->ps_epilog.states = shader->key.ps.epilog;
+ key->ps_epilog.states = shader->key.part.ps.epilog;
+}
+
+/**
+ * Build the GS prolog function. Rotate the input vertices for triangle strips
+ * with adjacency.
+ */
+static void si_build_gs_prolog_function(struct si_shader_context *ctx,
+ union si_shader_part_key *key)
+{
+ const unsigned num_sgprs = SI_GS_NUM_USER_SGPR + 2;
+ const unsigned num_vgprs = 8;
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMTypeRef params[32];
+ LLVMTypeRef returns[32];
+ LLVMValueRef func, ret;
+
+ for (unsigned i = 0; i < num_sgprs; ++i) {
+ params[i] = ctx->i32;
+ returns[i] = ctx->i32;
+ }
+
+ for (unsigned i = 0; i < num_vgprs; ++i) {
+ params[num_sgprs + i] = ctx->i32;
+ returns[num_sgprs + i] = ctx->f32;
+ }
+
+ /* Create the function. */
+ si_create_function(ctx, "gs_prolog", returns, num_sgprs + num_vgprs,
+ params, num_sgprs + num_vgprs, num_sgprs - 1);
+ func = ctx->main_fn;
+
+ /* Copy inputs to outputs. This should be no-op, as the registers match,
+ * but it will prevent the compiler from overwriting them unintentionally.
+ */
+ ret = ctx->return_value;
+ for (unsigned i = 0; i < num_sgprs; i++) {
+ LLVMValueRef p = LLVMGetParam(func, i);
+ ret = LLVMBuildInsertValue(builder, ret, p, i, "");
+ }
+ for (unsigned i = 0; i < num_vgprs; i++) {
+ LLVMValueRef p = LLVMGetParam(func, num_sgprs + i);
+ p = LLVMBuildBitCast(builder, p, ctx->f32, "");
+ ret = LLVMBuildInsertValue(builder, ret, p, num_sgprs + i, "");
+ }
+
+ if (key->gs_prolog.states.tri_strip_adj_fix) {
+ /* Remap the input vertices for every other primitive. */
+ const unsigned vtx_params[6] = {
+ num_sgprs,
+ num_sgprs + 1,
+ num_sgprs + 3,
+ num_sgprs + 4,
+ num_sgprs + 5,
+ num_sgprs + 6
+ };
+ LLVMValueRef prim_id, rotate;
+
+ prim_id = LLVMGetParam(func, num_sgprs + 2);
+ rotate = LLVMBuildTrunc(builder, prim_id, ctx->i1, "");
+
+ for (unsigned i = 0; i < 6; ++i) {
+ LLVMValueRef base, rotated, actual;
+ base = LLVMGetParam(func, vtx_params[i]);
+ rotated = LLVMGetParam(func, vtx_params[(i + 4) % 6]);
+ actual = LLVMBuildSelect(builder, rotate, rotated, base, "");
+ actual = LLVMBuildBitCast(builder, actual, ctx->f32, "");
+ ret = LLVMBuildInsertValue(builder, ret, actual, vtx_params[i], "");
+ }
+ }
+
+ LLVMBuildRet(builder, ret);
}
/**
LLVMValueRef out[48];
LLVMTypeRef function_type;
unsigned num_params;
- unsigned num_out_sgpr, num_out;
+ unsigned num_out;
+ MAYBE_UNUSED unsigned num_out_sgpr; /* used in debug checks */
unsigned num_sgprs, num_vgprs;
unsigned last_sgpr_param;
unsigned gprs;
for (unsigned i = 0; i < num_parts; ++i) {
- LLVMAddFunctionAttr(parts[i], LLVMAlwaysInlineAttribute);
+ lp_add_function_attr(parts[i], -1, LP_FUNC_ATTR_ALWAYSINLINE);
LLVMSetLinkage(parts[i], LLVMPrivateLinkage);
}
is_sgpr = ac_is_sgpr_param(param);
if (is_sgpr) {
+#if HAVE_LLVM < 0x0400
LLVMRemoveAttribute(param, LLVMByValAttribute);
- LLVMAddAttribute(param, LLVMInRegAttribute);
+#else
+ unsigned kind_id = LLVMGetEnumAttributeKindForName("byval", 5);
+ LLVMRemoveEnumAttributeAtIndex(parts[part], param_idx + 1, kind_id);
+#endif
+ lp_add_function_attr(parts[part], param_idx + 1, LP_FUNC_ATTR_INREG);
}
assert(out_idx + param_size <= (is_sgpr ? num_out_sgpr : num_out));
}
si_init_shader_ctx(&ctx, sscreen, shader, tm);
- ctx.no_prolog = is_monolithic;
- ctx.no_epilog = is_monolithic;
ctx.separate_prolog = !is_monolithic;
- if (ctx.type == PIPE_SHADER_VERTEX ||
- ctx.type == PIPE_SHADER_TESS_EVAL ||
- ctx.type == PIPE_SHADER_FRAGMENT) {
- ctx.no_prolog = false;
- ctx.no_epilog = false;
- }
-
- memset(shader->info.vs_output_param_offset, 0xff,
+ memset(shader->info.vs_output_param_offset, EXP_PARAM_UNDEFINED,
sizeof(shader->info.vs_output_param_offset));
shader->info.uses_instanceid = sel->info.uses_instanceid;
bool need_epilog;
need_prolog = sel->info.num_inputs;
- need_epilog = !shader->key.vs.as_es && !shader->key.vs.as_ls;
+ need_epilog = !shader->key.as_es && !shader->key.as_ls;
parts[need_prolog ? 1 : 0] = ctx.main_fn;
if (need_epilog) {
union si_shader_part_key epilog_key;
- si_get_vs_epilog_key(shader, &shader->key.vs.epilog, &epilog_key);
+ si_get_vs_epilog_key(shader, &shader->key.part.vs.epilog, &epilog_key);
si_build_vs_epilog_function(&ctx, &epilog_key);
parts[need_prolog ? 2 : 1] = ctx.main_fn;
}
si_build_wrapper_function(&ctx, parts, 1 + need_prolog + need_epilog,
need_prolog ? 1 : 0);
+ } else if (is_monolithic && ctx.type == PIPE_SHADER_TESS_CTRL) {
+ LLVMValueRef parts[2];
+ union si_shader_part_key epilog_key;
+
+ parts[0] = ctx.main_fn;
+
+ memset(&epilog_key, 0, sizeof(epilog_key));
+ epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
+ si_build_tcs_epilog_function(&ctx, &epilog_key);
+ parts[1] = ctx.main_fn;
+
+ si_build_wrapper_function(&ctx, parts, 2, 0);
} else if (is_monolithic && ctx.type == PIPE_SHADER_TESS_EVAL &&
- !shader->key.tes.as_es) {
+ !shader->key.as_es) {
LLVMValueRef parts[2];
union si_shader_part_key epilog_key;
parts[0] = ctx.main_fn;
- si_get_vs_epilog_key(shader, &shader->key.tes.epilog, &epilog_key);
+ si_get_vs_epilog_key(shader, &shader->key.part.tes.epilog, &epilog_key);
si_build_vs_epilog_function(&ctx, &epilog_key);
parts[1] = ctx.main_fn;
si_build_wrapper_function(&ctx, parts, 2, 0);
+ } else if (is_monolithic && ctx.type == PIPE_SHADER_GEOMETRY) {
+ LLVMValueRef parts[2];
+ union si_shader_part_key prolog_key;
+
+ parts[1] = ctx.main_fn;
+
+ memset(&prolog_key, 0, sizeof(prolog_key));
+ prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
+ si_build_gs_prolog_function(&ctx, &prolog_key);
+ parts[0] = ctx.main_fn;
+
+ si_build_wrapper_function(&ctx, parts, 2, 1);
} else if (is_monolithic && ctx.type == PIPE_SHADER_FRAGMENT) {
LLVMValueRef parts[3];
union si_shader_part_key prolog_key;
si_llvm_finalize_module(&ctx,
r600_extra_shader_checks(&sscreen->b, ctx.type));
- /* Post-optimization transformations. */
+ /* Post-optimization transformations and analysis. */
si_eliminate_const_vs_outputs(&ctx);
+ if ((debug && debug->debug_message) ||
+ r600_can_dump_shader(&sscreen->b, ctx.type))
+ si_count_scratch_private_memory(&ctx);
+
/* Compile to bytecode. */
r = si_compile_llvm(sscreen, &shader->binary, &shader->config, tm,
mod, debug, ctx.type, "TGSI shader");
* LLVM 3.9svn has this bug.
*/
if (sel->type == PIPE_SHADER_COMPUTE) {
- unsigned *props = sel->info.properties;
unsigned wave_size = 64;
unsigned max_vgprs = 256;
unsigned max_sgprs = sscreen->b.chip_class >= VI ? 800 : 512;
unsigned max_sgprs_per_wave = 128;
- unsigned max_block_threads;
-
- if (props[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH])
- max_block_threads = props[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
- props[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
- props[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
- else
- max_block_threads = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
-
+ unsigned max_block_threads = si_get_max_workgroup_size(shader);
unsigned min_waves_per_cu = DIV_ROUND_UP(max_block_threads, wave_size);
unsigned min_waves_per_simd = DIV_ROUND_UP(min_waves_per_cu, 4);
shader->info.num_input_vgprs += 1;
}
- if (ctx.type == PIPE_SHADER_GEOMETRY) {
- shader->gs_copy_shader = CALLOC_STRUCT(si_shader);
- shader->gs_copy_shader->selector = shader->selector;
- ctx.shader = shader->gs_copy_shader;
- r = si_generate_gs_copy_shader(sscreen, &ctx,
- shader, debug);
- if (r) {
- free(shader->gs_copy_shader);
- shader->gs_copy_shader = NULL;
- return r;
- }
- }
-
return 0;
}
*
* \param sscreen screen
* \param list list of shader parts of the same category
+ * \param type shader type
* \param key shader part key
+ * \param prolog whether the part being requested is a prolog
* \param tm LLVM target machine
* \param debug debug callback
- * \param compile the callback responsible for compilation
+ * \param build the callback responsible for building the main function
* \return non-NULL on success
*/
static struct si_shader_part *
si_get_shader_part(struct si_screen *sscreen,
struct si_shader_part **list,
+ enum pipe_shader_type type,
+ bool prolog,
union si_shader_part_key *key,
LLVMTargetMachineRef tm,
struct pipe_debug_callback *debug,
- bool (*compile)(struct si_screen *,
- LLVMTargetMachineRef,
- struct pipe_debug_callback *,
- struct si_shader_part *))
+ void (*build)(struct si_shader_context *,
+ union si_shader_part_key *),
+ const char *name)
{
struct si_shader_part *result;
/* Compile a new one. */
result = CALLOC_STRUCT(si_shader_part);
result->key = *key;
- if (!compile(sscreen, tm, debug, result)) {
+
+ struct si_shader shader = {};
+ struct si_shader_context ctx;
+ struct gallivm_state *gallivm = &ctx.gallivm;
+
+ si_init_shader_ctx(&ctx, sscreen, &shader, tm);
+ ctx.type = type;
+
+ switch (type) {
+ case PIPE_SHADER_VERTEX:
+ break;
+ case PIPE_SHADER_TESS_CTRL:
+ assert(!prolog);
+ shader.key.part.tcs.epilog = key->tcs_epilog.states;
+ break;
+ case PIPE_SHADER_GEOMETRY:
+ assert(prolog);
+ break;
+ case PIPE_SHADER_FRAGMENT:
+ if (prolog)
+ shader.key.part.ps.prolog = key->ps_prolog.states;
+ else
+ shader.key.part.ps.epilog = key->ps_epilog.states;
+ break;
+ default:
+ unreachable("bad shader part");
+ }
+
+ build(&ctx, key);
+
+ /* Compile. */
+ si_llvm_finalize_module(&ctx,
+ r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_FRAGMENT));
+
+ if (si_compile_llvm(sscreen, &result->binary, &result->config, tm,
+ gallivm->module, debug, ctx.type, name)) {
FREE(result);
- pipe_mutex_unlock(sscreen->shader_parts_mutex);
- return NULL;
+ result = NULL;
+ goto out;
}
result->next = *list;
*list = result;
+
+out:
+ si_llvm_dispose(&ctx);
pipe_mutex_unlock(sscreen->shader_parts_mutex);
return result;
}
si_llvm_build_ret(ctx, ret);
}
-/**
- * Create a vertex shader prolog.
- */
-static bool si_compile_vs_prolog(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
- struct pipe_debug_callback *debug,
- struct si_shader_part *out)
-{
- union si_shader_part_key *key = &out->key;
- struct si_shader shader = {};
- struct si_shader_context ctx;
- struct gallivm_state *gallivm = &ctx.gallivm;
- bool status = true;
-
- si_init_shader_ctx(&ctx, sscreen, &shader, tm);
- ctx.type = PIPE_SHADER_VERTEX;
-
- si_build_vs_prolog_function(&ctx, key);
-
- /* Compile. */
- si_llvm_finalize_module(&ctx,
- r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_VERTEX));
-
- if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
- gallivm->module, debug, ctx.type,
- "Vertex Shader Prolog"))
- status = false;
-
- si_llvm_dispose(&ctx);
- return status;
-}
-
/**
* Build the vertex shader epilog function. This is also used by the tessellation
* evaluation shader compiled as VS.
LLVMBuildRetVoid(gallivm->builder);
}
-/**
- * Compile the vertex shader epilog. This is also used by the tessellation
- * evaluation shader compiled as VS.
- */
-static bool si_compile_vs_epilog(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
- struct pipe_debug_callback *debug,
- struct si_shader_part *out)
-{
- union si_shader_part_key *key = &out->key;
- struct si_shader_context ctx;
- struct gallivm_state *gallivm = &ctx.gallivm;
- bool status = true;
-
- si_init_shader_ctx(&ctx, sscreen, NULL, tm);
- ctx.type = PIPE_SHADER_VERTEX;
-
- si_build_vs_epilog_function(&ctx, key);
-
- /* Compile. */
- si_llvm_finalize_module(&ctx,
- r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_VERTEX));
-
- if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
- gallivm->module, debug, ctx.type,
- "Vertex Shader Epilog"))
- status = false;
-
- si_llvm_dispose(&ctx);
- return status;
-}
-
/**
* Create & compile a vertex shader epilog. This a helper used by VS and TES.
*/
si_get_vs_epilog_key(shader, states, &epilog_key);
shader->epilog = si_get_shader_part(sscreen, &sscreen->vs_epilogs,
+ PIPE_SHADER_VERTEX, true,
&epilog_key, tm, debug,
- si_compile_vs_epilog);
+ si_build_vs_epilog_function,
+ "Vertex Shader Epilog");
return shader->epilog != NULL;
}
if (info->num_inputs) {
shader->prolog =
si_get_shader_part(sscreen, &sscreen->vs_prologs,
+ PIPE_SHADER_VERTEX, true,
&prolog_key, tm, debug,
- si_compile_vs_prolog);
+ si_build_vs_prolog_function,
+ "Vertex Shader Prolog");
if (!shader->prolog)
return false;
}
/* Get the epilog. */
- if (!shader->key.vs.as_es && !shader->key.vs.as_ls &&
+ if (!shader->key.as_es && !shader->key.as_ls &&
!si_get_vs_epilog(sscreen, tm, shader, debug,
- &shader->key.vs.epilog))
+ &shader->key.part.vs.epilog))
return false;
return true;
struct si_shader *shader,
struct pipe_debug_callback *debug)
{
- if (shader->key.tes.as_es)
+ if (shader->key.as_es)
return true;
/* TES compiled as VS. */
return si_get_vs_epilog(sscreen, tm, shader, debug,
- &shader->key.tes.epilog);
+ &shader->key.part.tes.epilog);
}
/**
- * Compile the TCS epilog. This writes tesselation factors to memory based on
- * the output primitive type of the tesselator (determined by TES).
+ * Compile the TCS epilog function. This writes tesselation factors to memory
+ * based on the output primitive type of the tesselator (determined by TES).
*/
-static bool si_compile_tcs_epilog(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
- struct pipe_debug_callback *debug,
- struct si_shader_part *out)
+static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
+ union si_shader_part_key *key)
{
- union si_shader_part_key *key = &out->key;
- struct si_shader shader = {};
- struct si_shader_context ctx;
- struct gallivm_state *gallivm = &ctx.gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx.soa.bld_base;
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
LLVMTypeRef params[16];
LLVMValueRef func;
int last_sgpr, num_params;
- bool status = true;
-
- si_init_shader_ctx(&ctx, sscreen, &shader, tm);
- ctx.type = PIPE_SHADER_TESS_CTRL;
- shader.key.tcs.epilog = key->tcs_epilog.states;
/* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
- params[SI_PARAM_RW_BUFFERS] = const_array(ctx.v16i8, SI_NUM_RW_BUFFERS);
- params[SI_PARAM_CONST_BUFFERS] = ctx.i64;
- params[SI_PARAM_SAMPLERS] = ctx.i64;
- params[SI_PARAM_IMAGES] = ctx.i64;
- params[SI_PARAM_SHADER_BUFFERS] = ctx.i64;
- params[SI_PARAM_TCS_OFFCHIP_LAYOUT] = ctx.i32;
- params[SI_PARAM_TCS_OUT_OFFSETS] = ctx.i32;
- params[SI_PARAM_TCS_OUT_LAYOUT] = ctx.i32;
- params[SI_PARAM_TCS_IN_LAYOUT] = ctx.i32;
- params[ctx.param_oc_lds = SI_PARAM_TCS_OC_LDS] = ctx.i32;
- params[SI_PARAM_TESS_FACTOR_OFFSET] = ctx.i32;
+ params[SI_PARAM_RW_BUFFERS] = const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
+ params[SI_PARAM_CONST_BUFFERS] = ctx->i64;
+ params[SI_PARAM_SAMPLERS] = ctx->i64;
+ params[SI_PARAM_IMAGES] = ctx->i64;
+ params[SI_PARAM_SHADER_BUFFERS] = ctx->i64;
+ params[SI_PARAM_TCS_OFFCHIP_LAYOUT] = ctx->i32;
+ params[SI_PARAM_TCS_OUT_OFFSETS] = ctx->i32;
+ params[SI_PARAM_TCS_OUT_LAYOUT] = ctx->i32;
+ params[SI_PARAM_TCS_IN_LAYOUT] = ctx->i32;
+ params[ctx->param_oc_lds = SI_PARAM_TCS_OC_LDS] = ctx->i32;
+ params[SI_PARAM_TESS_FACTOR_OFFSET] = ctx->i32;
last_sgpr = SI_PARAM_TESS_FACTOR_OFFSET;
num_params = last_sgpr + 1;
- params[num_params++] = ctx.i32; /* patch index within the wave (REL_PATCH_ID) */
- params[num_params++] = ctx.i32; /* invocation ID within the patch */
- params[num_params++] = ctx.i32; /* LDS offset where tess factors should be loaded from */
+ params[num_params++] = ctx->i32; /* patch index within the wave (REL_PATCH_ID) */
+ params[num_params++] = ctx->i32; /* invocation ID within the patch */
+ params[num_params++] = ctx->i32; /* LDS offset where tess factors should be loaded from */
/* Create the function. */
- si_create_function(&ctx, "tcs_epilog", NULL, 0, params, num_params, last_sgpr);
- declare_tess_lds(&ctx);
- func = ctx.main_fn;
+ si_create_function(ctx, "tcs_epilog", NULL, 0, params, num_params, last_sgpr);
+ declare_tess_lds(ctx);
+ func = ctx->main_fn;
si_write_tess_factors(bld_base,
LLVMGetParam(func, last_sgpr + 1),
LLVMGetParam(func, last_sgpr + 2),
LLVMGetParam(func, last_sgpr + 3));
- /* Compile. */
LLVMBuildRetVoid(gallivm->builder);
- si_llvm_finalize_module(&ctx,
- r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_TESS_CTRL));
-
- if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
- gallivm->module, debug, ctx.type,
- "Tessellation Control Shader Epilog"))
- status = false;
-
- si_llvm_dispose(&ctx);
- return status;
}
/**
/* Get the epilog. */
memset(&epilog_key, 0, sizeof(epilog_key));
- epilog_key.tcs_epilog.states = shader->key.tcs.epilog;
+ epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
shader->epilog = si_get_shader_part(sscreen, &sscreen->tcs_epilogs,
+ PIPE_SHADER_TESS_CTRL, false,
&epilog_key, tm, debug,
- si_compile_tcs_epilog);
+ si_build_tcs_epilog_function,
+ "Tessellation Control Shader Epilog");
return shader->epilog != NULL;
}
+/**
+ * Select and compile (or reuse) GS parts (prolog).
+ */
+static bool si_shader_select_gs_parts(struct si_screen *sscreen,
+ LLVMTargetMachineRef tm,
+ struct si_shader *shader,
+ struct pipe_debug_callback *debug)
+{
+ union si_shader_part_key prolog_key;
+
+ if (!shader->key.part.gs.prolog.tri_strip_adj_fix)
+ return true;
+
+ memset(&prolog_key, 0, sizeof(prolog_key));
+ prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
+
+ shader->prolog = si_get_shader_part(sscreen, &sscreen->gs_prologs,
+ PIPE_SHADER_GEOMETRY, true,
+ &prolog_key, tm, debug,
+ si_build_gs_prolog_function,
+ "Geometry Shader Prolog");
+ return shader->prolog != NULL;
+}
+
/**
* Build the pixel shader prolog function. This handles:
* - two-side color selection and interpolation
interp[1] = LLVMBuildExtractValue(gallivm->builder, ret,
interp_vgpr + 1, "");
interp_ij = lp_build_gather_values(gallivm, interp, 2);
- interp_ij = LLVMBuildBitCast(gallivm->builder, interp_ij,
- ctx->v2i32, "");
}
/* Use the absolute location of the input. */
si_llvm_build_ret(ctx, ret);
}
-/**
- * Compile the pixel shader prolog.
- */
-static bool si_compile_ps_prolog(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
- struct pipe_debug_callback *debug,
- struct si_shader_part *out)
-{
- union si_shader_part_key *key = &out->key;
- struct si_shader shader = {};
- struct si_shader_context ctx;
- struct gallivm_state *gallivm = &ctx.gallivm;
- bool status = true;
-
- si_init_shader_ctx(&ctx, sscreen, &shader, tm);
- ctx.type = PIPE_SHADER_FRAGMENT;
- shader.key.ps.prolog = key->ps_prolog.states;
-
- si_build_ps_prolog_function(&ctx, key);
-
- /* Compile. */
- si_llvm_finalize_module(&ctx,
- r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_FRAGMENT));
-
- if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
- gallivm->module, debug, ctx.type,
- "Fragment Shader Prolog"))
- status = false;
-
- si_llvm_dispose(&ctx);
- return status;
-}
-
/**
* Build the pixel shader epilog function. This handles everything that must be
* emulated for pixel shader exports. (alpha-test, format conversions, etc)
LLVMBuildRetVoid(gallivm->builder);
}
-
-/**
- * Compile the pixel shader epilog to a binary for concatenation.
- */
-static bool si_compile_ps_epilog(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
- struct pipe_debug_callback *debug,
- struct si_shader_part *out)
-{
- union si_shader_part_key *key = &out->key;
- struct si_shader shader = {};
- struct si_shader_context ctx;
- struct gallivm_state *gallivm = &ctx.gallivm;
- bool status = true;
-
- si_init_shader_ctx(&ctx, sscreen, &shader, tm);
- ctx.type = PIPE_SHADER_FRAGMENT;
- shader.key.ps.epilog = key->ps_epilog.states;
-
- si_build_ps_epilog_function(&ctx, key);
-
- /* Compile. */
- si_llvm_finalize_module(&ctx,
- r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_FRAGMENT));
-
- if (si_compile_llvm(sscreen, &out->binary, &out->config, tm,
- gallivm->module, debug, ctx.type,
- "Fragment Shader Epilog"))
- status = false;
-
- si_llvm_dispose(&ctx);
- return status;
-}
-
/**
* Select and compile (or reuse) pixel shader parts (prolog & epilog).
*/
if (si_need_ps_prolog(&prolog_key)) {
shader->prolog =
si_get_shader_part(sscreen, &sscreen->ps_prologs,
+ PIPE_SHADER_FRAGMENT, true,
&prolog_key, tm, debug,
- si_compile_ps_prolog);
+ si_build_ps_prolog_function,
+ "Fragment Shader Prolog");
if (!shader->prolog)
return false;
}
shader->epilog =
si_get_shader_part(sscreen, &sscreen->ps_epilogs,
+ PIPE_SHADER_FRAGMENT, false,
&epilog_key, tm, debug,
- si_compile_ps_epilog);
+ si_build_ps_epilog_function,
+ "Fragment Shader Epilog");
if (!shader->epilog)
return false;
/* Enable POS_FIXED_PT if polygon stippling is enabled. */
- if (shader->key.ps.prolog.poly_stipple) {
+ if (shader->key.part.ps.prolog.poly_stipple) {
shader->config.spi_ps_input_ena |= S_0286CC_POS_FIXED_PT_ENA(1);
assert(G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr));
}
/* Set up the enable bits for per-sample shading if needed. */
- if (shader->key.ps.prolog.force_persp_sample_interp &&
+ if (shader->key.part.ps.prolog.force_persp_sample_interp &&
(G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_ena) ||
G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTER_ENA;
shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
}
- if (shader->key.ps.prolog.force_linear_sample_interp &&
+ if (shader->key.part.ps.prolog.force_linear_sample_interp &&
(G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_ena) ||
G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTER_ENA;
shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
}
- if (shader->key.ps.prolog.force_persp_center_interp &&
+ if (shader->key.part.ps.prolog.force_persp_center_interp &&
(G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
shader->config.spi_ps_input_ena &= C_0286CC_PERSP_SAMPLE_ENA;
shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
}
- if (shader->key.ps.prolog.force_linear_center_interp &&
+ if (shader->key.part.ps.prolog.force_linear_center_interp &&
(G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_SAMPLE_ENA;
/* The sample mask input is always enabled, because the API shader always
* passes it through to the epilog. Disable it here if it's unused.
*/
- if (!shader->key.ps.epilog.poly_line_smoothing &&
+ if (!shader->key.part.ps.epilog.poly_line_smoothing &&
!shader->selector->info.reads_samplemask)
shader->config.spi_ps_input_ena &= C_0286CC_SAMPLE_COVERAGE_ENA;
return true;
}
-static void si_fix_num_sgprs(struct si_shader *shader)
+void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
+ unsigned *lds_size)
+{
+ /* SPI barrier management bug:
+ * Make sure we have at least 4k of LDS in use to avoid the bug.
+ * It applies to workgroup sizes of more than one wavefront.
+ */
+ if (sscreen->b.family == CHIP_BONAIRE ||
+ sscreen->b.family == CHIP_KABINI ||
+ sscreen->b.family == CHIP_MULLINS)
+ *lds_size = MAX2(*lds_size, 8);
+}
+
+static void si_fix_resource_usage(struct si_screen *sscreen,
+ struct si_shader *shader)
{
unsigned min_sgprs = shader->info.num_input_sgprs + 2; /* VCC */
shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
+
+ if (shader->selector->type == PIPE_SHADER_COMPUTE &&
+ si_get_max_workgroup_size(shader) > 64) {
+ si_multiwave_lds_size_workaround(sscreen,
+ &shader->config.lds_size);
+ }
}
int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
/* LS, ES, VS are compiled on demand if the main part hasn't been
* compiled for that stage.
+ *
+ * Vertex shaders are compiled on demand when a vertex fetch
+ * workaround must be applied.
*/
- if (!mainp ||
- (sel->type == PIPE_SHADER_VERTEX &&
- (shader->key.vs.as_es != mainp->key.vs.as_es ||
- shader->key.vs.as_ls != mainp->key.vs.as_ls)) ||
- (sel->type == PIPE_SHADER_TESS_EVAL &&
- shader->key.tes.as_es != mainp->key.tes.as_es) ||
- (sel->type == PIPE_SHADER_TESS_CTRL &&
- shader->key.tcs.epilog.inputs_to_copy) ||
- sel->type == PIPE_SHADER_COMPUTE) {
+ if (shader->is_monolithic) {
/* Monolithic shader (compiled as a whole, has many variants,
* may take a long time to compile).
*/
if (!si_shader_select_tes_parts(sscreen, tm, shader, debug))
return -1;
break;
+ case PIPE_SHADER_GEOMETRY:
+ if (!si_shader_select_gs_parts(sscreen, tm, shader, debug))
+ return -1;
+ break;
case PIPE_SHADER_FRAGMENT:
if (!si_shader_select_ps_parts(sscreen, tm, shader, debug))
return -1;
}
}
- si_fix_num_sgprs(shader);
+ si_fix_resource_usage(sscreen, shader);
si_shader_dump(sscreen, shader, debug, sel->info.processor,
stderr);
void si_shader_destroy(struct si_shader *shader)
{
- if (shader->gs_copy_shader) {
- si_shader_destroy(shader->gs_copy_shader);
- FREE(shader->gs_copy_shader);
- }
-
if (shader->scratch_bo)
r600_resource_reference(&shader->scratch_bo, NULL);