LOCAL_ADDR_SPACE = 3,
};
-#define SENDMSG_GS 2
-#define SENDMSG_GS_DONE 3
-
-#define SENDMSG_GS_OP_NOP (0 << 4)
-#define SENDMSG_GS_OP_CUT (1 << 4)
-#define SENDMSG_GS_OP_EMIT (2 << 4)
-#define SENDMSG_GS_OP_EMIT_CUT (3 << 4)
-
/**
* Returns a unique index for a semantic name and index. The index must be
* less than 64, so that a 64-bit bitmask of used inputs or outputs can be
LLVMGetParam(radeon_bld->main_fn, param_start_instance), "");
}
+/* Bitcast <4 x float> to <2 x double>, extract the component, and convert
+ * to float. */
+static LLVMValueRef extract_double_to_float(struct si_shader_context *ctx,
+ LLVMValueRef vec4,
+ unsigned double_index)
+{
+ LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMTypeRef f64 = LLVMDoubleTypeInContext(ctx->gallivm.context);
+ LLVMValueRef dvec2 = LLVMBuildBitCast(builder, vec4,
+ LLVMVectorType(f64, 2), "");
+ LLVMValueRef index = LLVMConstInt(ctx->i32, double_index, 0);
+ LLVMValueRef value = LLVMBuildExtractElement(builder, dvec2, index, "");
+ return LLVMBuildFPTrunc(builder, value, ctx->f32, "");
+}
+
static void declare_input_vs(
struct si_shader_context *ctx,
unsigned input_index,
unsigned chan;
unsigned fix_fetch;
+ unsigned num_fetches;
+ unsigned fetch_stride;
LLVMValueRef t_list_ptr;
LLVMValueRef t_offset;
LLVMValueRef t_list;
- LLVMValueRef attribute_offset;
- LLVMValueRef buffer_index;
+ LLVMValueRef vertex_index;
LLVMValueRef args[3];
- LLVMValueRef input;
+ LLVMValueRef input[3];
/* Load the T list */
t_list_ptr = LLVMGetParam(ctx->main_fn, SI_PARAM_VERTEX_BUFFERS);
t_list = ac_build_indexed_load_const(&ctx->ac, t_list_ptr, t_offset);
- /* Build the attribute offset */
- attribute_offset = lp_build_const_int32(gallivm, 0);
-
- buffer_index = LLVMGetParam(ctx->main_fn,
+ vertex_index = LLVMGetParam(ctx->main_fn,
ctx->param_vertex_index0 +
input_index);
+ fix_fetch = ctx->shader->key.mono.vs.fix_fetch[input_index];
+
+ /* Do multiple loads for special formats. */
+ switch (fix_fetch) {
+ case SI_FIX_FETCH_RGB_64_FLOAT:
+ num_fetches = 3; /* 3 2-dword loads */
+ fetch_stride = 8;
+ break;
+ case SI_FIX_FETCH_RGBA_64_FLOAT:
+ num_fetches = 2; /* 2 4-dword loads */
+ fetch_stride = 16;
+ break;
+ case SI_FIX_FETCH_RGB_8:
+ case SI_FIX_FETCH_RGB_8_INT:
+ num_fetches = 3;
+ fetch_stride = 1;
+ break;
+ case SI_FIX_FETCH_RGB_16:
+ case SI_FIX_FETCH_RGB_16_INT:
+ num_fetches = 3;
+ fetch_stride = 2;
+ break;
+ default:
+ num_fetches = 1;
+ fetch_stride = 0;
+ }
+
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,
- LP_FUNC_ATTR_READNONE);
+ args[2] = vertex_index;
+
+ for (unsigned i = 0; i < num_fetches; i++) {
+ args[1] = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
+
+ input[i] = lp_build_intrinsic(gallivm->builder,
+ "llvm.SI.vs.load.input", ctx->v4f32, args, 3,
+ LP_FUNC_ATTR_READNONE);
+ }
/* Break up the vec4 into individual components */
for (chan = 0; chan < 4; chan++) {
LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
out[chan] = LLVMBuildExtractElement(gallivm->builder,
- input, llvm_chan, "");
+ input[0], llvm_chan, "");
}
- fix_fetch = (ctx->shader->key.mono.vs.fix_fetch >> (4 * input_index)) & 0xf;
-
switch (fix_fetch) {
case SI_FIX_FETCH_A2_SNORM:
case SI_FIX_FETCH_A2_SSCALED:
out[chan], ctx->f32, "");
}
break;
+ case SI_FIX_FETCH_RG_64_FLOAT:
+ for (chan = 0; chan < 2; chan++)
+ out[chan] = extract_double_to_float(ctx, input[0], chan);
+
+ out[2] = LLVMConstReal(ctx->f32, 0);
+ out[3] = LLVMConstReal(ctx->f32, 1);
+ break;
+ case SI_FIX_FETCH_RGB_64_FLOAT:
+ for (chan = 0; chan < 3; chan++)
+ out[chan] = extract_double_to_float(ctx, input[chan], 0);
+
+ out[3] = LLVMConstReal(ctx->f32, 1);
+ break;
+ case SI_FIX_FETCH_RGBA_64_FLOAT:
+ for (chan = 0; chan < 4; chan++) {
+ out[chan] = extract_double_to_float(ctx, input[chan / 2],
+ chan % 2);
+ }
+ break;
+ case SI_FIX_FETCH_RGB_8:
+ case SI_FIX_FETCH_RGB_8_INT:
+ case SI_FIX_FETCH_RGB_16:
+ case SI_FIX_FETCH_RGB_16_INT:
+ for (chan = 0; chan < 3; chan++) {
+ out[chan] = LLVMBuildExtractElement(gallivm->builder,
+ input[chan],
+ ctx->i32_0, "");
+ }
+ if (fix_fetch == SI_FIX_FETCH_RGB_8 ||
+ fix_fetch == SI_FIX_FETCH_RGB_16) {
+ out[3] = LLVMConstReal(ctx->f32, 1);
+ } else {
+ out[3] = LLVMBuildBitCast(gallivm->builder, ctx->i32_1,
+ ctx->f32, "");
+ }
+ break;
}
}
static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context *bld_base)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- LLVMValueRef args[2];
- args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_NOP | SENDMSG_GS_DONE);
- args[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID);
- lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
- ctx->voidt, args, 2, 0);
+ ac_emit_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
+ LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
}
static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context *bld_base)
struct lp_build_context *uint_bld = &bld_base->uint_bld;
struct lp_build_emit_data txf_emit_data = *emit_data;
LLVMValueRef txf_address[4];
- unsigned txf_count = count;
+ /* We only need .xy for non-arrays, and .xyz for arrays. */
+ unsigned txf_count = target == TGSI_TEXTURE_2D_MSAA ? 2 : 3;
struct tgsi_full_instruction inst = {};
memcpy(txf_address, address, sizeof(txf_address));
- if (target == TGSI_TEXTURE_2D_MSAA) {
- txf_address[2] = bld_base->uint_bld.zero;
- }
- txf_address[3] = bld_base->uint_bld.zero;
-
/* Read FMASK using TXF. */
inst.Instruction.Opcode = TGSI_OPCODE_TXF;
inst.Texture.Texture = target;
txf_emit_data.output[0],
uint_bld->zero, "");
- unsigned sample_chan = target == TGSI_TEXTURE_2D_MSAA ? 2 : 3;
+ unsigned sample_chan = txf_count; /* the sample index is last */
LLVMValueRef sample_index4 =
LLVMBuildMul(gallivm->builder, address[sample_chan], four, "");
emit_data->output[emit_data->chan] = samples;
}
-/*
- * SI implements derivatives using the local data store (LDS)
- * All writes to the LDS happen in all executing threads at
- * the same time. TID is the Thread ID for the current
- * thread and is a value between 0 and 63, representing
- * the thread's position in the wavefront.
- *
- * For the pixel shader threads are grouped into quads of four pixels.
- * The TIDs of the pixels of a quad are:
- *
- * +------+------+
- * |4n + 0|4n + 1|
- * +------+------+
- * |4n + 2|4n + 3|
- * +------+------+
- *
- * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
- * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
- * the current pixel's column, and masking with 0xfffffffe yields the TID
- * of the left pixel of the current pixel's row.
- *
- * Adding 1 yields the TID of the pixel to the right of the left pixel, and
- * adding 2 yields the TID of the pixel below the top pixel.
- */
-/* masks for thread ID. */
-#define TID_MASK_TOP_LEFT 0xfffffffc
-#define TID_MASK_TOP 0xfffffffd
-#define TID_MASK_LEFT 0xfffffffe
-
static void si_llvm_emit_ddxy(
const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct si_shader_context *ctx = si_shader_context(bld_base);
struct gallivm_state *gallivm = bld_base->base.gallivm;
unsigned opcode = emit_data->info->opcode;
- LLVMValueRef thread_id, tl, trbl, tl_tid, trbl_tid, val, args[2];
+ LLVMValueRef val;
int idx;
unsigned mask;
- thread_id = ac_get_thread_id(&ctx->ac);
-
if (opcode == TGSI_OPCODE_DDX_FINE)
- mask = TID_MASK_LEFT;
+ mask = AC_TID_MASK_LEFT;
else if (opcode == TGSI_OPCODE_DDY_FINE)
- mask = TID_MASK_TOP;
+ mask = AC_TID_MASK_TOP;
else
- mask = TID_MASK_TOP_LEFT;
-
- tl_tid = LLVMBuildAnd(gallivm->builder, thread_id,
- lp_build_const_int32(gallivm, mask), "");
+ mask = AC_TID_MASK_TOP_LEFT;
/* for DDX we want to next X pixel, DDY next Y pixel. */
idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2;
- trbl_tid = LLVMBuildAdd(gallivm->builder, tl_tid,
- lp_build_const_int32(gallivm, idx), "");
val = LLVMBuildBitCast(gallivm->builder, emit_data->args[0], ctx->i32, "");
-
- if (ctx->screen->has_ds_bpermute) {
- args[0] = LLVMBuildMul(gallivm->builder, tl_tid,
- lp_build_const_int32(gallivm, 4), "");
- args[1] = val;
- tl = lp_build_intrinsic(gallivm->builder,
- "llvm.amdgcn.ds.bpermute", ctx->i32,
- 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, LP_FUNC_ATTR_READNONE);
- } else {
- LLVMValueRef store_ptr, load_ptr0, load_ptr1;
-
- store_ptr = ac_build_gep0(&ctx->ac, ctx->lds, thread_id);
- load_ptr0 = ac_build_gep0(&ctx->ac, ctx->lds, tl_tid);
- load_ptr1 = ac_build_gep0(&ctx->ac, ctx->lds, trbl_tid);
-
- LLVMBuildStore(gallivm->builder, val, store_ptr);
- tl = LLVMBuildLoad(gallivm->builder, load_ptr0, "");
- trbl = LLVMBuildLoad(gallivm->builder, load_ptr1, "");
- }
-
- tl = LLVMBuildBitCast(gallivm->builder, tl, ctx->f32, "");
- trbl = LLVMBuildBitCast(gallivm->builder, trbl, ctx->f32, "");
-
- emit_data->output[emit_data->chan] =
- LLVMBuildFSub(gallivm->builder, trbl, tl, "");
+ val = ac_emit_ddxy(&ctx->ac, ctx->screen->has_ds_bpermute,
+ mask, idx, ctx->lds, val);
+ emit_data->output[emit_data->chan] = val;
}
/*
SI_PARAM_GS2VS_OFFSET);
LLVMValueRef gs_next_vertex;
LLVMValueRef can_emit, kill;
- LLVMValueRef args[2];
unsigned chan, offset;
int i;
unsigned stream;
LLVMBuildStore(gallivm->builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
/* Signal vertex emission */
- args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_EMIT | SENDMSG_GS | (stream << 8));
- args[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID);
- lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
- ctx->voidt, args, 2, 0);
-
+ ac_emit_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
+ LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
if (!use_kill)
lp_build_endif(&if_state);
}
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;
- LLVMValueRef args[2];
unsigned stream;
/* Signal primitive cut */
stream = si_llvm_get_stream(bld_base, emit_data);
- args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_CUT | SENDMSG_GS | (stream << 8));
- args[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID);
- lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
- ctx->voidt, args, 2, 0);
+ ac_emit_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
+ LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
}
static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
r600_resource_reference(&shader->bo, NULL);
shader->bo = (struct r600_resource*)
pipe_buffer_create(&sscreen->b.b, 0,
- PIPE_USAGE_IMMUTABLE, bo_size);
+ PIPE_USAGE_IMMUTABLE,
+ align(bo_size, SI_CPDMA_ALIGNMENT));
if (!shader->bo)
return -ENOMEM;
conf->spilled_vgprs, conf->private_mem_vgprs);
}
-static const char *si_get_shader_name(struct si_shader *shader,
- unsigned processor)
+const char *si_get_shader_name(struct si_shader *shader, unsigned processor)
{
switch (processor) {
case PIPE_SHADER_VERTEX:
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%"PRIx64"\n", key->mono.vs.fix_fetch);
+
+ fprintf(f, " mono.vs.fix_fetch = {");
+ for (i = 0; i < SI_MAX_ATTRIBS; i++)
+ fprintf(f, !i ? "%u" : ", %u", key->mono.vs.fix_fetch[i]);
+ fprintf(f, "}\n");
break;
case PIPE_SHADER_TESS_CTRL:
struct pipe_debug_callback *debug)
{
struct si_shader_selector *sel = shader->selector;
- struct si_shader *mainp = sel->main_shader_part;
+ struct si_shader *mainp = *si_get_main_shader_part(sel, &shader->key);
int r;
/* LS, ES, VS are compiled on demand if the main part hasn't been