#include "gallivm/lp_bld_intr.h"
#include "gallivm/lp_bld_logic.h"
#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_bitarit.h"
#include "gallivm/lp_bld_flow.h"
#include "radeon/r600_cs.h"
#include "radeon/radeon_llvm.h"
LLVMValueRef samplers[SI_NUM_SAMPLER_STATES];
LLVMValueRef so_buffers[4];
LLVMValueRef esgs_ring;
- LLVMValueRef gsvs_ring;
- LLVMValueRef gs_next_vertex;
+ LLVMValueRef gsvs_ring[4];
+ LLVMValueRef gs_next_vertex[4];
};
static struct si_shader_context * si_shader_context(
args[0] = t_list;
args[1] = attribute_offset;
args[2] = buffer_index;
- input = build_intrinsic(gallivm->builder,
+ input = lp_build_intrinsic(gallivm->builder,
"llvm.SI.vs.load.input", vec4_type, args, 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[8] = uint->zero; /* TFE */
return LLVMBuildBitCast(gallivm->builder,
- build_intrinsic(gallivm->builder,
+ lp_build_intrinsic(gallivm->builder,
"llvm.SI.buffer.load.dword.i32.i32",
i32, args, 9,
LLVMReadOnlyAttribute | LLVMNoUnwindAttribute),
args[0] = llvm_chan;
args[1] = attr_number;
- front = build_intrinsic(gallivm->builder, intr_name,
+ front = lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[1] = back_attr_number;
- back = build_intrinsic(gallivm->builder, intr_name,
+ back = lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[2] = params;
args[3] = interp_param;
radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 0)] =
- build_intrinsic(gallivm->builder, intr_name,
+ lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 1)] =
args[2] = params;
args[3] = interp_param;
radeon_bld->inputs[soa_index] =
- build_intrinsic(gallivm->builder, intr_name,
+ lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
{
LLVMValueRef args[2] = {resource, offset};
- return build_intrinsic(builder, "llvm.SI.load.const", return_type, args, 2,
+ return lp_build_intrinsic(builder, "llvm.SI.load.const", return_type, args, 2,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
args[0] = values[2 * chan];
args[1] = values[2 * chan + 1];
args[chan + 5] =
- build_intrinsic(base->gallivm->builder,
+ lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.packf16",
LLVMInt32TypeInContext(base->gallivm->context),
args, 2,
lp_build_const_float(gallivm, 1.0f),
lp_build_const_float(gallivm, -1.0f));
- build_intrinsic(gallivm->builder,
+ lp_build_intrinsic(gallivm->builder,
"llvm.AMDGPU.kill",
LLVMVoidTypeInContext(gallivm->context),
&arg, 1, 0);
} else {
- build_intrinsic(gallivm->builder,
+ lp_build_intrinsic(gallivm->builder,
"llvm.AMDGPU.kilp",
LLVMVoidTypeInContext(gallivm->context),
NULL, 0, 0);
SI_PARAM_SAMPLE_COVERAGE);
coverage = bitcast(bld_base, TGSI_TYPE_SIGNED, coverage);
- coverage = build_intrinsic(gallivm->builder, "llvm.ctpop.i32",
+ coverage = lp_build_intrinsic(gallivm->builder, "llvm.ctpop.i32",
bld_base->int_bld.elem_type,
&coverage, 1, LLVMReadNoneAttribute);
lp_build_intrinsic(gallivm->builder, name,
LLVMVoidTypeInContext(gallivm->context),
- args, Elements(args));
+ args, Elements(args), 0);
}
static void build_tbuffer_store_dwords(struct si_shader_context *shader,
LLVMValueRef so_vtx_count =
unpack_param(shader, shader->param_streamout_config, 16, 7);
- LLVMValueRef tid = build_intrinsic(builder, "llvm.SI.tid", i32,
+ LLVMValueRef tid = lp_build_intrinsic(builder, "llvm.SI.tid", i32,
NULL, 0, LLVMReadNoneAttribute);
/* can_emit = tid < so_vtx_count; */
LLVMValueRef can_emit =
LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
+ LLVMValueRef stream_id =
+ unpack_param(shader, shader->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. */
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)
break;
}
+ 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(shader, shader->so_buffers[buf_idx],
vdata, num_comps,
so_write_offset[buf_idx],
LLVMConstInt(i32, 0, 0),
so->output[i].dst_offset*4);
+ lp_build_endif(&if_ctx_stream);
}
}
lp_build_endif(&if_ctx);
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
}
if (semantic_name == TGSI_SEMANTIC_CLIPDIST) {
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- pos_args[i], 9);
+ pos_args[i], 9, 0);
}
}
args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_NOP | SENDMSG_GS_DONE);
args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID);
- build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
+ lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
LLVMVoidTypeInContext(gallivm->context), args, 2,
LLVMNoUnwindAttribute);
}
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- last_args, 9);
+ last_args, 9, 0);
}
/* This instruction will be emitted at the end of the shader. */
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
}
}
} else {
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
}
}
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
else
memcpy(last_args, args, sizeof(args));
}
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- last_args, 9);
+ last_args, 9, 0);
}
static void build_tex_intrinsic(const struct lp_build_tgsi_action * action,
const struct tgsi_full_instruction * inst = emit_data->inst;
unsigned opcode = inst->Instruction.Opcode;
unsigned target = inst->Texture.Texture;
- LLVMValueRef coords[5];
+ LLVMValueRef coords[5], derivs[6];
LLVMValueRef address[16];
int ref_pos;
unsigned num_coords = tgsi_util_get_texture_coord_dim(target, &ref_pos);
unsigned chan;
unsigned sampler_src;
unsigned sampler_index;
+ unsigned num_deriv_channels = 0;
bool has_offset = HAVE_LLVM >= 0x0305 ? inst->Texture.NumOffsets > 0 : false;
LLVMValueRef res_ptr, samp_ptr;
}
}
- if (target == TGSI_TEXTURE_CUBE ||
- target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
- radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
-
/* Pack user derivatives */
if (opcode == TGSI_OPCODE_TXD) {
- int num_deriv_channels, param;
+ int param, num_src_deriv_channels;
switch (target) {
case TGSI_TEXTURE_3D:
+ num_src_deriv_channels = 3;
num_deriv_channels = 3;
break;
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_SHADOWRECT:
case TGSI_TEXTURE_2D_ARRAY:
case TGSI_TEXTURE_SHADOW2D_ARRAY:
+ num_src_deriv_channels = 2;
+ num_deriv_channels = 2;
+ break;
case TGSI_TEXTURE_CUBE:
case TGSI_TEXTURE_SHADOWCUBE:
case TGSI_TEXTURE_CUBE_ARRAY:
case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
+ /* Cube derivatives will be converted to 2D. */
+ num_src_deriv_channels = 3;
num_deriv_channels = 2;
break;
case TGSI_TEXTURE_1D:
case TGSI_TEXTURE_SHADOW1D:
case TGSI_TEXTURE_1D_ARRAY:
case TGSI_TEXTURE_SHADOW1D_ARRAY:
+ num_src_deriv_channels = 1;
num_deriv_channels = 1;
break;
default:
assert(0); /* no other targets are valid here */
}
- for (param = 1; param <= 2; param++)
- for (chan = 0; chan < num_deriv_channels; chan++)
- address[count++] = lp_build_emit_fetch(bld_base, inst, param, chan);
+ for (param = 0; param < 2; param++)
+ for (chan = 0; chan < num_src_deriv_channels; chan++)
+ derivs[param * num_src_deriv_channels + chan] =
+ lp_build_emit_fetch(bld_base, inst, param+1, chan);
}
+ if (target == TGSI_TEXTURE_CUBE ||
+ target == TGSI_TEXTURE_CUBE_ARRAY ||
+ target == TGSI_TEXTURE_SHADOWCUBE ||
+ target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
+ radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords, derivs);
+
+ if (opcode == TGSI_OPCODE_TXD)
+ for (int i = 0; i < num_deriv_channels * 2; i++)
+ address[count++] = derivs[i];
+
/* Pack texture coordinates */
address[count++] = coords[0];
if (num_coords > 1)
emit_data->inst->Texture.NumOffsets > 0 : false;
if (target == TGSI_TEXTURE_BUFFER) {
- emit_data->output[emit_data->chan] = build_intrinsic(
+ emit_data->output[emit_data->chan] = lp_build_intrinsic(
base->gallivm->builder,
"llvm.SI.vs.load.input", emit_data->dst_type,
emit_data->args, emit_data->arg_count,
is_shadow ? ".c" : "", infix, has_offset ? ".o" : "",
LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
- emit_data->output[emit_data->chan] = build_intrinsic(
+ 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 | LLVMNoUnwindAttribute);
sprintf(intr_name, "%s.v%ui32", name,
LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
- emit_data->output[emit_data->chan] = build_intrinsic(
+ 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 | LLVMNoUnwindAttribute);
}
}
+/*
+ * 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,
LLVMTypeRef i32;
unsigned swizzle[4];
unsigned c;
+ int idx;
+ unsigned mask;
i32 = LLVMInt32TypeInContext(gallivm->context);
indices[0] = bld_base->uint_bld.zero;
- indices[1] = build_intrinsic(gallivm->builder, "llvm.SI.tid", i32,
+ indices[1] = lp_build_intrinsic(gallivm->builder, "llvm.SI.tid", i32,
NULL, 0, LLVMReadNoneAttribute);
store_ptr = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
indices, 2, "");
+ if (opcode == TGSI_OPCODE_DDX_FINE)
+ mask = TID_MASK_LEFT;
+ else if (opcode == TGSI_OPCODE_DDY_FINE)
+ mask = TID_MASK_TOP;
+ else
+ mask = TID_MASK_TOP_LEFT;
+
indices[1] = LLVMBuildAnd(gallivm->builder, indices[1],
- lp_build_const_int32(gallivm, 0xfffffffc), "");
+ lp_build_const_int32(gallivm, mask), "");
load_ptr0 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
indices, 2, "");
+ /* for DDX we want to next X pixel, DDY next Y pixel. */
+ idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2;
indices[1] = LLVMBuildAdd(gallivm->builder, indices[1],
- lp_build_const_int32(gallivm,
- opcode == TGSI_OPCODE_DDX ? 1 : 2),
- "");
+ lp_build_const_int32(gallivm, idx), "");
load_ptr1 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
indices, 2, "");
emit_data->output[0] = lp_build_gather_values(gallivm, result, 4);
}
+/*
+ * this takes an I,J coordinate pair,
+ * and works out the X and Y derivatives.
+ * it returns DDX(I), DDX(J), DDY(I), DDY(J).
+ */
+static LLVMValueRef si_llvm_emit_ddxy_interp(
+ struct lp_build_tgsi_context *bld_base,
+ LLVMValueRef interp_ij)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct lp_build_context *base = &bld_base->base;
+ LLVMValueRef indices[2];
+ LLVMValueRef store_ptr, load_ptr_x, load_ptr_y, load_ptr_ddx, load_ptr_ddy, temp, temp2;
+ LLVMValueRef tl, tr, bl, result[4];
+ LLVMTypeRef i32;
+ unsigned c;
+
+ i32 = LLVMInt32TypeInContext(gallivm->context);
+
+ indices[0] = bld_base->uint_bld.zero;
+ indices[1] = lp_build_intrinsic(gallivm->builder, "llvm.SI.tid", i32,
+ NULL, 0, LLVMReadNoneAttribute);
+ store_ptr = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ temp = LLVMBuildAnd(gallivm->builder, indices[1],
+ lp_build_const_int32(gallivm, TID_MASK_LEFT), "");
+
+ temp2 = LLVMBuildAnd(gallivm->builder, indices[1],
+ lp_build_const_int32(gallivm, TID_MASK_TOP), "");
+
+ indices[1] = temp;
+ load_ptr_x = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ indices[1] = temp2;
+ load_ptr_y = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ indices[1] = LLVMBuildAdd(gallivm->builder, temp,
+ lp_build_const_int32(gallivm, 1), "");
+ load_ptr_ddx = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ indices[1] = LLVMBuildAdd(gallivm->builder, temp2,
+ lp_build_const_int32(gallivm, 2), "");
+ load_ptr_ddy = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ for (c = 0; c < 2; ++c) {
+ LLVMValueRef store_val;
+ LLVMValueRef c_ll = lp_build_const_int32(gallivm, c);
+
+ store_val = LLVMBuildExtractElement(gallivm->builder,
+ interp_ij, c_ll, "");
+ LLVMBuildStore(gallivm->builder,
+ store_val,
+ store_ptr);
+
+ tl = LLVMBuildLoad(gallivm->builder, load_ptr_x, "");
+ tl = LLVMBuildBitCast(gallivm->builder, tl, base->elem_type, "");
+
+ tr = LLVMBuildLoad(gallivm->builder, load_ptr_ddx, "");
+ tr = LLVMBuildBitCast(gallivm->builder, tr, base->elem_type, "");
+
+ result[c] = LLVMBuildFSub(gallivm->builder, tr, tl, "");
+
+ tl = LLVMBuildLoad(gallivm->builder, load_ptr_y, "");
+ tl = LLVMBuildBitCast(gallivm->builder, tl, base->elem_type, "");
+
+ bl = LLVMBuildLoad(gallivm->builder, load_ptr_ddy, "");
+ bl = LLVMBuildBitCast(gallivm->builder, bl, base->elem_type, "");
+
+ result[c + 2] = LLVMBuildFSub(gallivm->builder, bl, tl, "");
+ }
+
+ return lp_build_gather_values(gallivm, result, 4);
+}
+
+static void interp_fetch_args(
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ const struct tgsi_full_instruction *inst = emit_data->inst;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET) {
+ /* offset is in second src, first two channels */
+ emit_data->args[0] = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 1,
+ 0);
+ emit_data->args[1] = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 1,
+ 1);
+ emit_data->arg_count = 2;
+ } else if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
+ LLVMValueRef sample_position;
+ LLVMValueRef sample_id;
+ LLVMValueRef halfval = lp_build_const_float(gallivm, 0.5f);
+
+ /* fetch sample ID, then fetch its sample position,
+ * and place into first two channels.
+ */
+ sample_id = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 1, 0);
+ sample_id = LLVMBuildBitCast(gallivm->builder, sample_id,
+ LLVMInt32TypeInContext(gallivm->context),
+ "");
+ sample_position = load_sample_position(&si_shader_ctx->radeon_bld, sample_id);
+
+ emit_data->args[0] = LLVMBuildExtractElement(gallivm->builder,
+ sample_position,
+ lp_build_const_int32(gallivm, 0), "");
+
+ emit_data->args[0] = LLVMBuildFSub(gallivm->builder, emit_data->args[0], halfval, "");
+ emit_data->args[1] = LLVMBuildExtractElement(gallivm->builder,
+ sample_position,
+ lp_build_const_int32(gallivm, 1), "");
+ emit_data->args[1] = LLVMBuildFSub(gallivm->builder, emit_data->args[1], halfval, "");
+ emit_data->arg_count = 2;
+ }
+}
+
+static void build_interp_intrinsic(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct si_shader *shader = si_shader_ctx->shader;
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMValueRef interp_param;
+ const struct tgsi_full_instruction *inst = emit_data->inst;
+ const char *intr_name;
+ int input_index;
+ int chan;
+ int i;
+ LLVMValueRef attr_number;
+ LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
+ LLVMValueRef params = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_PRIM_MASK);
+ int interp_param_idx;
+ unsigned location;
+
+ assert(inst->Src[0].Register.File == TGSI_FILE_INPUT);
+ input_index = inst->Src[0].Register.Index;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
+ inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE)
+ location = TGSI_INTERPOLATE_LOC_CENTER;
+ else
+ location = TGSI_INTERPOLATE_LOC_CENTROID;
+
+ interp_param_idx = lookup_interp_param_index(shader->ps_input_interpolate[input_index],
+ location);
+ if (interp_param_idx == -1)
+ return;
+ else if (interp_param_idx)
+ interp_param = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, interp_param_idx);
+ else
+ interp_param = NULL;
+
+ attr_number = lp_build_const_int32(gallivm,
+ shader->ps_input_param_offset[input_index]);
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
+ inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
+ LLVMValueRef ij_out[2];
+ LLVMValueRef ddxy_out = si_llvm_emit_ddxy_interp(bld_base, interp_param);
+
+ /*
+ * take the I then J parameters, and the DDX/Y for it, and
+ * calculate the IJ inputs for the interpolator.
+ * temp1 = ddx * offset/sample.x + I;
+ * interp_param.I = ddy * offset/sample.y + temp1;
+ * temp1 = ddx * offset/sample.x + J;
+ * interp_param.J = ddy * offset/sample.y + temp1;
+ */
+ for (i = 0; i < 2; i++) {
+ LLVMValueRef ix_ll = lp_build_const_int32(gallivm, i);
+ LLVMValueRef iy_ll = lp_build_const_int32(gallivm, i + 2);
+ LLVMValueRef ddx_el = LLVMBuildExtractElement(gallivm->builder,
+ ddxy_out, ix_ll, "");
+ LLVMValueRef ddy_el = LLVMBuildExtractElement(gallivm->builder,
+ ddxy_out, iy_ll, "");
+ LLVMValueRef interp_el = LLVMBuildExtractElement(gallivm->builder,
+ interp_param, ix_ll, "");
+ LLVMValueRef temp1, temp2;
+
+ interp_el = LLVMBuildBitCast(gallivm->builder, interp_el,
+ LLVMFloatTypeInContext(gallivm->context), "");
+
+ temp1 = LLVMBuildFMul(gallivm->builder, ddx_el, emit_data->args[0], "");
+
+ temp1 = LLVMBuildFAdd(gallivm->builder, temp1, interp_el, "");
+
+ temp2 = LLVMBuildFMul(gallivm->builder, ddy_el, emit_data->args[1], "");
+
+ temp2 = LLVMBuildFAdd(gallivm->builder, temp2, temp1, "");
+
+ ij_out[i] = LLVMBuildBitCast(gallivm->builder,
+ temp2,
+ LLVMIntTypeInContext(gallivm->context, 32), "");
+ }
+ 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 < 2; 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,
+ input_type, args, args[3] ? 4 : 3,
+ LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+ }
+}
+
+static unsigned si_llvm_get_stream(struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates;
+ struct tgsi_src_register src0 = emit_data->inst->Src[0].Register;
+ unsigned stream;
+
+ assert(src0.File == TGSI_FILE_IMMEDIATE);
+
+ stream = LLVMConstIntGetZExtValue(imms[src0.Index][src0.SwizzleX]) & 0x3;
+ return stream;
+}
+
/* Emit one vertex from the geometry shader */
static void si_llvm_emit_vertex(
const struct lp_build_tgsi_action *action,
LLVMValueRef args[2];
unsigned chan;
int i;
+ unsigned stream;
+
+ stream = si_llvm_get_stream(bld_base, emit_data);
/* Write vertex attribute values to GSVS ring */
- gs_next_vertex = LLVMBuildLoad(gallivm->builder, si_shader_ctx->gs_next_vertex, "");
+ gs_next_vertex = LLVMBuildLoad(gallivm->builder,
+ si_shader_ctx->gs_next_vertex[stream],
+ "");
/* If this thread has already emitted the declared maximum number of
* vertices, kill it: excessive vertex emissions are not supposed to
kill = lp_build_select(&bld_base->base, can_emit,
lp_build_const_float(gallivm, 1.0f),
lp_build_const_float(gallivm, -1.0f));
- build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
- LLVMVoidTypeInContext(gallivm->context), &kill, 1, 0);
+
+ lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
+ LLVMVoidTypeInContext(gallivm->context), &kill, 1, 0);
for (i = 0; i < info->num_outputs; i++) {
LLVMValueRef *out_ptr =
out_val = LLVMBuildBitCast(gallivm->builder, out_val, i32, "");
build_tbuffer_store(si_shader_ctx,
- si_shader_ctx->gsvs_ring,
+ si_shader_ctx->gsvs_ring[stream],
out_val, 1,
voffset, soffset, 0,
V_008F0C_BUF_DATA_FORMAT_32,
}
gs_next_vertex = lp_build_add(uint, gs_next_vertex,
lp_build_const_int32(gallivm, 1));
- LLVMBuildStore(gallivm->builder, gs_next_vertex, si_shader_ctx->gs_next_vertex);
+
+ LLVMBuildStore(gallivm->builder, gs_next_vertex, si_shader_ctx->gs_next_vertex[stream]);
/* Signal vertex emission */
- args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_EMIT | SENDMSG_GS);
+ args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_EMIT | SENDMSG_GS | (stream << 8));
args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID);
- build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
+ lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
LLVMVoidTypeInContext(gallivm->context), args, 2,
LLVMNoUnwindAttribute);
}
struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMValueRef args[2];
+ unsigned stream;
/* Signal primitive cut */
- args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_CUT | SENDMSG_GS);
+ 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(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID);
- build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
+ lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
LLVMVoidTypeInContext(gallivm->context), args, 2,
LLVMNoUnwindAttribute);
}
{
struct gallivm_state *gallivm = bld_base->base.gallivm;
- build_intrinsic(gallivm->builder, "llvm.AMDGPU.barrier.local",
+ lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.barrier.local",
LLVMVoidTypeInContext(gallivm->context), NULL, 0,
LLVMNoUnwindAttribute);
}
.intr_name = "llvm.SI.resinfo"
};
+static const struct lp_build_tgsi_action interp_action = {
+ .fetch_args = interp_fetch_args,
+ .emit = build_interp_intrinsic,
+};
+
static void create_meta_data(struct si_shader_context *si_shader_ctx)
{
struct gallivm_state *gallivm = si_shader_ctx->radeon_bld.soa.bld_base.base.gallivm;
if (bld_base->info &&
(bld_base->info->opcode_count[TGSI_OPCODE_DDX] > 0 ||
- bld_base->info->opcode_count[TGSI_OPCODE_DDY] > 0))
+ bld_base->info->opcode_count[TGSI_OPCODE_DDY] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_DDX_FINE] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_DDY_FINE] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_INTERP_OFFSET] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_INTERP_SAMPLE] > 0))
si_shader_ctx->lds =
LLVMAddGlobalInAddressSpace(gallivm->module,
LLVMArrayType(i32, 64),
build_indexed_load_const(si_shader_ctx, buf_ptr, offset);
}
- if (si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY ||
- si_shader_ctx->shader->is_gs_copy_shader) {
+ if (si_shader_ctx->shader->is_gs_copy_shader) {
LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
- si_shader_ctx->gsvs_ring =
+ si_shader_ctx->gsvs_ring[0] =
build_indexed_load_const(si_shader_ctx, buf_ptr, offset);
}
+ if (si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY) {
+ int i;
+ for (i = 0; i < 4; i++) {
+ LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS + i);
+
+ si_shader_ctx->gsvs_ring[i] =
+ build_indexed_load_const(si_shader_ctx, buf_ptr, offset);
+ }
+ }
}
void si_shader_binary_read_config(const struct si_screen *sscreen,
si_shader_binary_upload(sscreen, shader);
if (dump) {
- if (!binary->disassembled) {
+ if (binary->disasm_string) {
+ fprintf(stderr, "\nShader Disassembly:\n\n");
+ fprintf(stderr, "%s\n", binary->disasm_string);
+ } else {
fprintf(stderr, "SI CODE:\n");
for (i = 0; i < binary->code_size; i+=4 ) {
fprintf(stderr, "@0x%x: %02x%02x%02x%02x\n", i, binary->code[i + 3],
if (shader->scratch_bytes_per_wave == 0) {
FREE(shader->binary.code);
FREE(shader->binary.relocs);
- memset(&shader->binary, 0, sizeof(shader->binary));
+ memset(&shader->binary, 0,
+ offsetof(struct radeon_shader_binary, disasm_string));
}
return r;
}
preload_streamout_buffers(si_shader_ctx);
preload_ring_buffers(si_shader_ctx);
- args[0] = si_shader_ctx->gsvs_ring;
+ args[0] = si_shader_ctx->gsvs_ring[0];
args[1] = lp_build_mul_imm(uint,
LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
si_shader_ctx->param_vertex_id),
outputs[i].values[chan] =
LLVMBuildBitCast(gallivm->builder,
- build_intrinsic(gallivm->builder,
+ lp_build_intrinsic(gallivm->builder,
"llvm.SI.buffer.load.dword.i32.i32",
LLVMInt32TypeInContext(gallivm->context),
args, 9,
bld_base->info = poly_stipple ? &stipple_shader_info : &sel->info;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
+ bld_base->op_actions[TGSI_OPCODE_INTERP_CENTROID] = interp_action;
+ bld_base->op_actions[TGSI_OPCODE_INTERP_SAMPLE] = interp_action;
+ bld_base->op_actions[TGSI_OPCODE_INTERP_OFFSET] = interp_action;
+
bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TEX2] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXB] = tex_action;
bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy;
bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy;
+ bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy;
+ bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy;
bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
preload_ring_buffers(&si_shader_ctx);
if (si_shader_ctx.type == TGSI_PROCESSOR_GEOMETRY) {
- si_shader_ctx.gs_next_vertex =
- lp_build_alloca(bld_base->base.gallivm,
- bld_base->uint_bld.elem_type, "");
+ int i;
+ for (i = 0; i < 4; i++) {
+ si_shader_ctx.gs_next_vertex[i] =
+ lp_build_alloca(bld_base->base.gallivm,
+ bld_base->uint_bld.elem_type, "");
+ }
}
if (!lp_build_tgsi_llvm(bld_base, tokens)) {
FREE(shader->binary.code);
FREE(shader->binary.relocs);
+ FREE(shader->binary.disasm_string);
}