+/*
+ * Copyright 2012 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors:
+ * Tom Stellard <thomas.stellard@amd.com>
+ * Michel Dänzer <michel.daenzer@amd.com>
+ * Christian König <christian.koenig@amd.com>
+ */
+
#include "gallivm/lp_bld_tgsi_action.h"
#include "gallivm/lp_bld_const.h"
+#include "gallivm/lp_bld_gather.h"
#include "gallivm/lp_bld_intr.h"
+#include "gallivm/lp_bld_logic.h"
#include "gallivm/lp_bld_tgsi.h"
#include "radeon_llvm.h"
#include "radeon_llvm_emit.h"
#include "radeonsi_pipe.h"
#include "radeonsi_shader.h"
+#include "si_state.h"
#include "sid.h"
#include <assert.h>
#include <errno.h>
#include <stdio.h>
-/*
-static ps_remap_inputs(
- struct tgsi_llvm_context * tl_ctx,
- unsigned tgsi_index,
- unsigned tgsi_chan)
-{
- :
-}
-
-struct si_input
-{
- struct list_head head;
- unsigned tgsi_index;
- unsigned tgsi_chan;
- unsigned order;
-};
-*/
-
-
struct si_shader_context
{
struct radeon_llvm_context radeon_bld;
struct tgsi_parse_context parse;
struct tgsi_token * tokens;
struct si_pipe_shader *shader;
+ struct si_shader_key key;
unsigned type; /* TGSI_PROCESSOR_* specifies the type of shader. */
-/* unsigned num_inputs; */
+ unsigned ninput_emitted;
/* struct list_head inputs; */
/* unsigned * input_mappings *//* From TGSI to SI hw */
/* struct tgsi_shader_info info;*/
#define CENTROID_OFSET 4
#define USE_SGPR_MAX_SUFFIX_LEN 5
+#define CONST_ADDR_SPACE 2
+#define USER_SGPR_ADDR_SPACE 8
enum sgpr_type {
+ SGPR_CONST_PTR_F32,
+ SGPR_CONST_PTR_V4I32,
+ SGPR_CONST_PTR_V8I32,
SGPR_I32,
- SGPR_I64,
- SGPR_PTR_V4I32,
- SGPR_PTR_V8I32
+ SGPR_I64
};
+/**
+ * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad
+ *
+ * @param offset The offset parameter specifies the number of
+ * elements to offset, not the number of bytes or dwords. An element is the
+ * the type pointed to by the base_ptr parameter (e.g. int is the element of
+ * an int* pointer)
+ *
+ * When LLVM lowers the load instruction, it will convert the element offset
+ * into a dword offset automatically.
+ *
+ */
+static LLVMValueRef build_indexed_load(
+ struct gallivm_state * gallivm,
+ LLVMValueRef base_ptr,
+ LLVMValueRef offset)
+{
+ LLVMValueRef computed_ptr = LLVMBuildGEP(
+ gallivm->builder, base_ptr, &offset, 1, "");
+
+ return LLVMBuildLoad(gallivm->builder, computed_ptr, "");
+}
+
+/**
+ * Load a value stored in one of the user SGPRs
+ *
+ * @param sgpr This is the sgpr to load the value from. If you need to load a
+ * value that is stored in consecutive SGPR registers (e.g. a 64-bit pointer),
+ * then you should pass the index of the first SGPR that holds the value. For
+ * example, if you want to load a pointer that is stored in SGPRs 2 and 3, then
+ * use pass 2 for the sgpr parameter.
+ *
+ * The value of the sgpr parameter must also be aligned to the width of the type
+ * being loaded, so that the sgpr parameter is divisible by the dword width of the
+ * type. For example, if the value being loaded is two dwords wide, then the sgpr
+ * parameter must be divisible by two.
+ */
static LLVMValueRef use_sgpr(
struct gallivm_state * gallivm,
enum sgpr_type type,
unsigned sgpr)
{
LLVMValueRef sgpr_index;
- LLVMValueRef sgpr_value;
LLVMTypeRef ret_type;
+ LLVMValueRef ptr;
sgpr_index = lp_build_const_int32(gallivm, sgpr);
- if (type == SGPR_I32) {
- ret_type = LLVMInt32TypeInContext(gallivm->context);
- return lp_build_intrinsic_unary(gallivm->builder,
- "llvm.SI.use.sgpr.i32",
- ret_type, sgpr_index);
- }
+ switch (type) {
+ case SGPR_CONST_PTR_F32:
+ assert(sgpr % 2 == 0);
+ ret_type = LLVMFloatTypeInContext(gallivm->context);
+ ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
+ break;
- ret_type = LLVMInt64TypeInContext(gallivm->context);
- sgpr_value = lp_build_intrinsic_unary(gallivm->builder,
- "llvm.SI.use.sgpr.i64",
- ret_type, sgpr_index);
+ case SGPR_I32:
+ ret_type = LLVMInt32TypeInContext(gallivm->context);
+ break;
- switch (type) {
case SGPR_I64:
- return sgpr_value;
- case SGPR_PTR_V4I32:
+ assert(sgpr % 2 == 0);
+ ret_type= LLVMInt64TypeInContext(gallivm->context);
+ break;
+
+ case SGPR_CONST_PTR_V4I32:
+ assert(sgpr % 2 == 0);
ret_type = LLVMInt32TypeInContext(gallivm->context);
ret_type = LLVMVectorType(ret_type, 4);
- ret_type = LLVMPointerType(ret_type,
- 0 /*XXX: Specify address space*/);
- return LLVMBuildIntToPtr(gallivm->builder, sgpr_value,
- ret_type, "");
- case SGPR_PTR_V8I32:
+ ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
+ break;
+
+ case SGPR_CONST_PTR_V8I32:
+ assert(sgpr % 2 == 0);
ret_type = LLVMInt32TypeInContext(gallivm->context);
ret_type = LLVMVectorType(ret_type, 8);
- ret_type = LLVMPointerType(ret_type,
- 0 /*XXX: Specify address space*/);
- return LLVMBuildIntToPtr(gallivm->builder, sgpr_value,
- ret_type, "");
+ ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
+ break;
+
default:
assert(!"Unsupported SGPR type in use_sgpr()");
return NULL;
}
+
+ ret_type = LLVMPointerType(ret_type, USER_SGPR_ADDR_SPACE);
+ ptr = LLVMBuildIntToPtr(gallivm->builder, sgpr_index, ret_type, "");
+ return LLVMBuildLoad(gallivm->builder, ptr, "");
}
static void declare_input_vs(
{
LLVMValueRef t_list_ptr;
LLVMValueRef t_offset;
+ LLVMValueRef t_list;
LLVMValueRef attribute_offset;
LLVMValueRef buffer_index_reg;
- LLVMValueRef args[4];
+ LLVMValueRef args[3];
LLVMTypeRef vec4_type;
LLVMValueRef input;
struct lp_build_context * uint = &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
struct lp_build_context * base = &si_shader_ctx->radeon_bld.soa.bld_base.base;
- struct r600_context *rctx = si_shader_ctx->rctx;
- struct pipe_vertex_element *velem = &rctx->vertex_elements->elements[input_index];
+ //struct pipe_vertex_element *velem = &rctx->vertex_elements->elements[input_index];
unsigned chan;
- /* XXX: Communicate with the rest of the driver about which SGPR the T#
- * list pointer is going to be stored in. Hard code to SGPR[6:7] for
- * now */
- t_list_ptr = use_sgpr(base->gallivm, SGPR_I64, 3);
+ /* Load the T list */
+ t_list_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_V4I32, SI_SGPR_VERTEX_BUFFER);
- t_offset = lp_build_const_int32(base->gallivm,
- 4 * velem->vertex_buffer_index);
- attribute_offset = lp_build_const_int32(base->gallivm, velem->src_offset);
+ t_offset = lp_build_const_int32(base->gallivm, input_index);
+
+ t_list = build_indexed_load(base->gallivm, t_list_ptr, t_offset);
+
+ /* Build the attribute offset */
+ attribute_offset = lp_build_const_int32(base->gallivm, 0);
/* Load the buffer index is always, which is always stored in VGPR0
* for Vertex Shaders */
- buffer_index_reg = lp_build_intrinsic(base->gallivm->builder,
- "llvm.SI.vs.load.buffer.index", uint->elem_type, NULL, 0);
+ buffer_index_reg = build_intrinsic(base->gallivm->builder,
+ "llvm.SI.vs.load.buffer.index", uint->elem_type, NULL, 0,
+ LLVMReadNoneAttribute);
vec4_type = LLVMVectorType(base->elem_type, 4);
- args[0] = t_list_ptr;
- args[1] = t_offset;
- args[2] = attribute_offset;
- args[3] = buffer_index_reg;
+ args[0] = t_list;
+ args[1] = attribute_offset;
+ args[2] = buffer_index_reg;
input = lp_build_intrinsic(base->gallivm->builder,
- "llvm.SI.vs.load.input", vec4_type, args, 4);
+ "llvm.SI.vs.load.input", vec4_type, args, 3);
/* Break up the vec4 into individual components */
for (chan = 0; chan < 4; chan++) {
{
const char * intr_name;
unsigned chan;
+ struct si_shader *shader = &si_shader_ctx->shader->shader;
struct lp_build_context * base =
&si_shader_ctx->radeon_bld.soa.bld_base.base;
struct gallivm_state * gallivm = base->gallivm;
+ LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
/* This value is:
* [15:0] NewPrimMask (Bit mask for each quad. It is set it the
* [32:16] ParamOffset
*
*/
- /* XXX: This register number must be identical to the S_00B02C_USER_SGPR
- * register field value
- */
- LLVMValueRef params = use_sgpr(base->gallivm, SGPR_I32, 6);
+ LLVMValueRef params = use_sgpr(base->gallivm, SGPR_I32, SI_PS_NUM_USER_SGPR);
+ LLVMValueRef attr_number;
+
+ if (decl->Semantic.Name == TGSI_SEMANTIC_POSITION) {
+ for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
+ LLVMValueRef args[1];
+ unsigned soa_index =
+ radeon_llvm_reg_index_soa(input_index, chan);
+ args[0] = lp_build_const_int32(gallivm, chan);
+ si_shader_ctx->radeon_bld.inputs[soa_index] =
+ build_intrinsic(base->gallivm->builder,
+ "llvm.SI.fs.read.pos", input_type,
+ args, 1, LLVMReadNoneAttribute);
+ }
+ return;
+ }
+
+ if (decl->Semantic.Name == TGSI_SEMANTIC_FACE) {
+ LLVMValueRef face, is_face_positive;
+
+ face = build_intrinsic(gallivm->builder,
+ "llvm.SI.fs.read.face",
+ input_type,
+ NULL, 0, LLVMReadNoneAttribute);
+ is_face_positive = LLVMBuildFCmp(gallivm->builder,
+ LLVMRealUGT, face,
+ lp_build_const_float(gallivm, 0.0f),
+ "");
+
+ si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 0)] =
+ LLVMBuildSelect(gallivm->builder,
+ is_face_positive,
+ lp_build_const_float(gallivm, 1.0f),
+ lp_build_const_float(gallivm, 0.0f),
+ "");
+ si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 1)] =
+ si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 2)] =
+ lp_build_const_float(gallivm, 0.0f);
+ si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 3)] =
+ lp_build_const_float(gallivm, 1.0f);
+ return;
+ }
- /* XXX: Is this the input_index? */
- LLVMValueRef attr_number = lp_build_const_int32(gallivm, input_index);
+ shader->input[input_index].param_offset = shader->ninterp++;
+ attr_number = lp_build_const_int32(gallivm,
+ shader->input[input_index].param_offset);
/* XXX: Handle all possible interpolation modes */
switch (decl->Interp.Interpolate) {
case TGSI_INTERPOLATE_COLOR:
- if (si_shader_ctx->rctx->rasterizer->flatshade) {
+ /* XXX: Flat shading hangs the GPU */
+ if (si_shader_ctx->rctx->queued.named.rasterizer &&
+ si_shader_ctx->rctx->queued.named.rasterizer->flatshade) {
+#if 0
intr_name = "llvm.SI.fs.interp.constant";
+#else
+ intr_name = "llvm.SI.fs.interp.linear.center";
+#endif
} else {
if (decl->Interp.Centroid)
intr_name = "llvm.SI.fs.interp.persp.centroid";
}
break;
case TGSI_INTERPOLATE_CONSTANT:
+ /* XXX: Flat shading hangs the GPU */
+#if 0
intr_name = "llvm.SI.fs.interp.constant";
break;
+#endif
case TGSI_INTERPOLATE_LINEAR:
if (decl->Interp.Centroid)
intr_name = "llvm.SI.fs.interp.linear.centroid";
return;
}
+ if (!si_shader_ctx->ninput_emitted++) {
+ /* Enable whole quad mode */
+ lp_build_intrinsic(gallivm->builder,
+ "llvm.SI.wqm",
+ LLVMVoidTypeInContext(gallivm->context),
+ NULL, 0);
+ }
+
/* XXX: Could there be more than TGSI_NUM_CHANNELS (4) ? */
- for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
+ if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
+ si_shader_ctx->key.color_two_side) {
LLVMValueRef args[3];
- LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
- unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
- LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
- args[0] = llvm_chan;
- args[1] = attr_number;
+ LLVMValueRef face, is_face_positive;
+ LLVMValueRef back_attr_number =
+ lp_build_const_int32(gallivm,
+ shader->input[input_index].param_offset + 1);
+
+ face = build_intrinsic(gallivm->builder,
+ "llvm.SI.fs.read.face",
+ input_type,
+ NULL, 0, LLVMReadNoneAttribute);
+ is_face_positive = LLVMBuildFCmp(gallivm->builder,
+ LLVMRealUGT, face,
+ lp_build_const_float(gallivm, 0.0f),
+ "");
+
args[2] = params;
- si_shader_ctx->radeon_bld.inputs[soa_index] =
- lp_build_intrinsic(gallivm->builder, intr_name,
- input_type, args, 3);
+ for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
+ LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
+ unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
+ LLVMValueRef front, back;
+
+ args[0] = llvm_chan;
+ args[1] = attr_number;
+ front = build_intrinsic(base->gallivm->builder, intr_name,
+ input_type, args, 3, LLVMReadOnlyAttribute);
+
+ args[1] = back_attr_number;
+ back = build_intrinsic(base->gallivm->builder, intr_name,
+ input_type, args, 3, LLVMReadOnlyAttribute);
+
+ si_shader_ctx->radeon_bld.inputs[soa_index] =
+ LLVMBuildSelect(gallivm->builder,
+ is_face_positive,
+ front,
+ back,
+ "");
+ }
+
+ shader->ninterp++;
+ } else {
+ for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
+ LLVMValueRef args[3];
+ LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
+ unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
+ args[0] = llvm_chan;
+ args[1] = attr_number;
+ args[2] = params;
+ si_shader_ctx->radeon_bld.inputs[soa_index] =
+ build_intrinsic(base->gallivm->builder, intr_name,
+ input_type, args, 3, LLVMReadOnlyAttribute);
+ }
}
}
unsigned swizzle)
{
struct lp_build_context * base = &bld_base->base;
+ unsigned idx;
LLVMValueRef const_ptr;
LLVMValueRef offset;
+ LLVMValueRef load;
+
+ /* currently not supported */
+ if (reg->Register.Indirect) {
+ assert(0);
+ load = lp_build_const_int32(base->gallivm, 0);
+ return bitcast(bld_base, type, load);
+ }
- /* XXX: Assume the pointer to the constant buffer is being stored in
- * SGPR[0:1] */
- const_ptr = use_sgpr(base->gallivm, SGPR_I64, 0);
+ const_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_F32, SI_SGPR_CONST);
/* XXX: This assumes that the constant buffer is not packed, so
* CONST[0].x will have an offset of 0 and CONST[1].x will have an
* offset of 4. */
- offset = lp_build_const_int32(base->gallivm,
- (reg->Register.Index * 4) + swizzle);
+ idx = (reg->Register.Index * 4) + swizzle;
- return lp_build_intrinsic_binary(base->gallivm->builder,
- "llvm.SI.load.const", base->elem_type, const_ptr, offset);
-}
+ /* index loads above 255 are currently not supported */
+ if (idx > 255) {
+ assert(0);
+ idx = 0;
+ }
+ offset = lp_build_const_int32(base->gallivm, idx);
+ load = build_indexed_load(base->gallivm, const_ptr, offset);
+ return bitcast(bld_base, type, load);
+}
-/* Declare some intrinsics with the correct attributes */
-static void si_llvm_emit_prologue(struct lp_build_tgsi_context * bld_base)
+/* Initialize arguments for the shader export intrinsic */
+static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
+ struct tgsi_full_declaration *d,
+ unsigned index,
+ unsigned target,
+ LLVMValueRef *args)
{
- LLVMValueRef function;
- struct gallivm_state * gallivm = bld_base->base.gallivm;
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct lp_build_context *uint =
+ &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
+ struct lp_build_context *base = &bld_base->base;
+ unsigned compressed = 0;
+ unsigned chan;
+
+ if (si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT) {
+ int cbuf = target - V_008DFC_SQ_EXP_MRT;
+
+ if (cbuf >= 0 && cbuf < 8) {
+ struct r600_context *rctx = si_shader_ctx->rctx;
+ compressed = (si_shader_ctx->key.export_16bpc >> cbuf) & 0x1;
+
+ if (compressed)
+ si_shader_ctx->shader->spi_shader_col_format |=
+ V_028714_SPI_SHADER_FP16_ABGR << (4 * cbuf);
+ else
+ si_shader_ctx->shader->spi_shader_col_format |=
+ V_028714_SPI_SHADER_32_ABGR << (4 * cbuf);
+ }
+ }
+
+ if (compressed) {
+ /* Pixel shader needs to pack output values before export */
+ for (chan = 0; chan < 2; chan++ ) {
+ LLVMValueRef *out_ptr =
+ si_shader_ctx->radeon_bld.soa.outputs[index];
+ args[0] = LLVMBuildLoad(base->gallivm->builder,
+ out_ptr[2 * chan], "");
+ args[1] = LLVMBuildLoad(base->gallivm->builder,
+ out_ptr[2 * chan + 1], "");
+ args[chan + 5] =
+ build_intrinsic(base->gallivm->builder,
+ "llvm.SI.packf16",
+ LLVMInt32TypeInContext(base->gallivm->context),
+ args, 2,
+ LLVMReadNoneAttribute);
+ args[chan + 7] = args[chan + 5] =
+ LLVMBuildBitCast(base->gallivm->builder,
+ args[chan + 5],
+ LLVMFloatTypeInContext(base->gallivm->context),
+ "");
+ }
+
+ /* Set COMPR flag */
+ args[4] = uint->one;
+ } else {
+ for (chan = 0; chan < 4; chan++ ) {
+ LLVMValueRef out_ptr =
+ si_shader_ctx->radeon_bld.soa.outputs[index][chan];
+ /* +5 because the first output value will be
+ * the 6th argument to the intrinsic. */
+ args[chan + 5] = LLVMBuildLoad(base->gallivm->builder,
+ out_ptr, "");
+ }
- LLVMTypeRef i64 = LLVMInt64TypeInContext(gallivm->context);
- LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
+ /* Clear COMPR flag */
+ args[4] = uint->zero;
+ }
- /* declare i32 @llvm.SI.use.sgpr.i32(i32) */
- function = lp_declare_intrinsic(gallivm->module, "llvm.SI.use.sgpr.i32",
- i32, &i32, 1);
- LLVMAddFunctionAttr(function, LLVMReadNoneAttribute);
+ /* XXX: This controls which components of the output
+ * registers actually get exported. (e.g bit 0 means export
+ * X component, bit 1 means export Y component, etc.) I'm
+ * hard coding this to 0xf for now. In the future, we might
+ * want to do something else. */
+ args[0] = lp_build_const_int32(base->gallivm, 0xf);
+
+ /* Specify whether the EXEC mask represents the valid mask */
+ args[1] = uint->zero;
- /* declare i64 @llvm.SI.use.sgpr.i64(i32) */
- function = lp_declare_intrinsic(gallivm->module, "llvm.SI.use.sgpr.i64",
- i64, &i32, 1);
- LLVMAddFunctionAttr(function, LLVMReadNoneAttribute);
+ /* Specify whether this is the last export */
+ args[2] = uint->zero;
+
+ /* Specify the target we are exporting */
+ args[3] = lp_build_const_int32(base->gallivm, target);
+
+ /* XXX: We probably need to keep track of the output
+ * values, so we know what we are passing to the next
+ * stage. */
+}
+
+static void si_llvm_emit_prologue(struct lp_build_tgsi_context *bld_base)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ lp_build_intrinsic_unary(gallivm->builder,
+ "llvm.AMDGPU.shader.type",
+ LLVMVoidTypeInContext(gallivm->context),
+ lp_build_const_int32(gallivm, si_shader_ctx->type));
+}
+
+
+static void si_alpha_test(struct lp_build_tgsi_context *bld_base,
+ unsigned index)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+
+ if (si_shader_ctx->key.alpha_func != PIPE_FUNC_NEVER) {
+ LLVMValueRef out_ptr = si_shader_ctx->radeon_bld.soa.outputs[index][3];
+ LLVMValueRef alpha_pass =
+ lp_build_cmp(&bld_base->base,
+ si_shader_ctx->key.alpha_func,
+ LLVMBuildLoad(gallivm->builder, out_ptr, ""),
+ lp_build_const_float(gallivm, si_shader_ctx->key.alpha_ref));
+ LLVMValueRef arg =
+ lp_build_select(&bld_base->base,
+ alpha_pass,
+ lp_build_const_float(gallivm, 1.0f),
+ lp_build_const_float(gallivm, -1.0f));
+
+ build_intrinsic(gallivm->builder,
+ "llvm.AMDGPU.kill",
+ LLVMVoidTypeInContext(gallivm->context),
+ &arg, 1, 0);
+ } else {
+ build_intrinsic(gallivm->builder,
+ "llvm.AMDGPU.kilp",
+ LLVMVoidTypeInContext(gallivm->context),
+ NULL, 0, 0);
+ }
}
/* XXX: This is partially implemented for VS only at this point. It is not complete */
static void si_llvm_emit_epilogue(struct lp_build_tgsi_context * bld_base)
{
struct si_shader_context * si_shader_ctx = si_shader_context(bld_base);
- struct r600_shader * shader = &si_shader_ctx->shader->shader;
+ struct si_shader * shader = &si_shader_ctx->shader->shader;
struct lp_build_context * base = &bld_base->base;
struct lp_build_context * uint =
&si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
struct tgsi_parse_context *parse = &si_shader_ctx->parse;
+ LLVMValueRef args[9];
LLVMValueRef last_args[9] = { 0 };
+ unsigned color_count = 0;
+ unsigned param_count = 0;
+ int depth_index = -1, stencil_index = -1;
while (!tgsi_parse_end_of_tokens(parse)) {
- /* XXX: component_bits controls which components of the output
- * registers actually get exported. (e.g bit 0 means export
- * X component, bit 1 means export Y component, etc.) I'm
- * hard coding this to 0xf for now. In the future, we might
- * want to do something else. */
- unsigned component_bits = 0xf;
- unsigned chan;
struct tgsi_full_declaration *d =
&parse->FullToken.FullDeclaration;
- LLVMValueRef args[9];
unsigned target;
unsigned index;
- unsigned color_count = 0;
- unsigned param_count = 0;
int i;
tgsi_parse_token(parse);
shader->input[i].sid = d->Semantic.Index;
shader->input[i].interpolate = d->Interp.Interpolate;
shader->input[i].centroid = d->Interp.Centroid;
- break;
+ continue;
+
case TGSI_FILE_OUTPUT:
i = shader->noutput++;
shader->output[i].name = d->Semantic.Name;
shader->output[i].sid = d->Semantic.Index;
shader->output[i].interpolate = d->Interp.Interpolate;
break;
- }
- if (d->Declaration.File != TGSI_FILE_OUTPUT)
+ default:
continue;
+ }
for (index = d->Range.First; index <= d->Range.Last; index++) {
- for (chan = 0; chan < 4; chan++ ) {
- LLVMValueRef out_ptr =
- si_shader_ctx->radeon_bld.soa.outputs
- [index][chan];
- /* +5 because the first output value will be
- * the 6th argument to the intrinsic. */
- args[chan + 5]= LLVMBuildLoad(
- base->gallivm->builder, out_ptr, "");
- }
-
- /* XXX: We probably need to keep track of the output
- * values, so we know what we are passing to the next
- * stage. */
-
/* Select the correct target */
switch(d->Semantic.Name) {
- case TGSI_SEMANTIC_POSITION:
+ case TGSI_SEMANTIC_PSIZE:
target = V_008DFC_SQ_EXP_POS;
break;
+ case TGSI_SEMANTIC_POSITION:
+ if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
+ target = V_008DFC_SQ_EXP_POS;
+ break;
+ } else {
+ depth_index = index;
+ continue;
+ }
+ case TGSI_SEMANTIC_STENCIL:
+ stencil_index = index;
+ continue;
case TGSI_SEMANTIC_COLOR:
if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
+ case TGSI_SEMANTIC_BCOLOR:
target = V_008DFC_SQ_EXP_PARAM + param_count;
+ shader->output[i].param_offset = param_count;
param_count++;
} else {
target = V_008DFC_SQ_EXP_MRT + color_count;
+ if (color_count == 0 &&
+ si_shader_ctx->key.alpha_func != PIPE_FUNC_ALWAYS)
+ si_alpha_test(bld_base, index);
+
color_count++;
}
break;
+ case TGSI_SEMANTIC_FOG:
case TGSI_SEMANTIC_GENERIC:
target = V_008DFC_SQ_EXP_PARAM + param_count;
+ shader->output[i].param_offset = param_count;
param_count++;
break;
default:
d->Semantic.Name);
}
- /* Specify which components to enable */
- args[0] = lp_build_const_int32(base->gallivm,
- component_bits);
-
- /* Specify whether the EXEC mask represents the valid mask */
- args[1] = lp_build_const_int32(base->gallivm, 0);
-
- /* Specify whether this is the last export */
- args[2] = lp_build_const_int32(base->gallivm, 0);
-
- /* Specify the target we are exporting */
- args[3] = lp_build_const_int32(base->gallivm, target);
-
- /* Set COMPR flag to zero to export data as 32-bit */
- args[4] = uint->zero;
+ si_llvm_init_export_args(bld_base, d, index, target, args);
if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX ?
(d->Semantic.Name == TGSI_SEMANTIC_POSITION) :
}
}
+ if (depth_index >= 0 || stencil_index >= 0) {
+ LLVMValueRef out_ptr;
+ unsigned mask = 0;
+
+ /* Specify the target we are exporting */
+ args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRTZ);
+
+ if (depth_index >= 0) {
+ out_ptr = si_shader_ctx->radeon_bld.soa.outputs[depth_index][2];
+ args[5] = LLVMBuildLoad(base->gallivm->builder, out_ptr, "");
+ mask |= 0x1;
+
+ if (stencil_index < 0) {
+ args[6] =
+ args[7] =
+ args[8] = args[5];
+ }
+ }
+
+ if (stencil_index >= 0) {
+ out_ptr = si_shader_ctx->radeon_bld.soa.outputs[stencil_index][1];
+ args[7] =
+ args[8] =
+ args[6] = LLVMBuildLoad(base->gallivm->builder, out_ptr, "");
+ mask |= 0x2;
+
+ if (depth_index < 0)
+ args[5] = args[6];
+ }
+
+ /* Specify which components to enable */
+ args[0] = lp_build_const_int32(base->gallivm, mask);
+
+ args[1] =
+ args[2] =
+ args[4] = uint->zero;
+
+ if (last_args[0])
+ lp_build_intrinsic(base->gallivm->builder,
+ "llvm.SI.export",
+ LLVMVoidTypeInContext(base->gallivm->context),
+ args, 9);
+ else
+ memcpy(last_args, args, sizeof(args));
+ }
+
+ if (!last_args[0]) {
+ assert(si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT);
+
+ /* Specify which components to enable */
+ last_args[0] = lp_build_const_int32(base->gallivm, 0x0);
+
+ /* Specify the target we are exporting */
+ last_args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRT);
+
+ /* Set COMPR flag to zero to export data as 32-bit */
+ last_args[4] = uint->zero;
+
+ /* dummy bits */
+ last_args[5]= uint->zero;
+ last_args[6]= uint->zero;
+ last_args[7]= uint->zero;
+ last_args[8]= uint->zero;
+
+ si_shader_ctx->shader->spi_shader_col_format |=
+ V_028714_SPI_SHADER_32_ABGR;
+ }
+
/* Specify whether the EXEC mask represents the valid mask */
last_args[1] = lp_build_const_int32(base->gallivm,
si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT);
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ const struct tgsi_full_instruction * inst = emit_data->inst;
+ LLVMValueRef ptr;
+ LLVMValueRef offset;
+ LLVMValueRef coords[5];
+ unsigned chan;
+
/* WriteMask */
- emit_data->args[0] = lp_build_const_int32(bld_base->base.gallivm,
- emit_data->inst->Dst[0].Register.WriteMask);
+ /* XXX: should be optimized using emit_data->inst->Dst[0].Register.WriteMask*/
+ emit_data->args[0] = lp_build_const_int32(bld_base->base.gallivm, 0xf);
/* Coordinates */
/* XXX: Not all sample instructions need 4 address arguments. */
- emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
- 0, LP_CHAN_ALL);
+ if (inst->Instruction.Opcode == TGSI_OPCODE_TXP)
+ coords[3] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W)
+;
+
+ for (chan = 0; chan < 3; chan++ ) {
+ coords[chan] = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 0,
+ chan);
+ if (inst->Instruction.Opcode == TGSI_OPCODE_TXP)
+ coords[chan] = lp_build_emit_llvm_binary(bld_base,
+ TGSI_OPCODE_DIV,
+ coords[chan],
+ coords[3]);
+ }
+
+ coords[3] = bld_base->base.one;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_TEX2 ||
+ inst->Instruction.Opcode == TGSI_OPCODE_TXB2 ||
+ inst->Instruction.Opcode == TGSI_OPCODE_TXL2) {
+ /* These instructions have additional operand that should be packed
+ * into the cube coord vector by radeon_llvm_emit_prepare_cube_coords.
+ * That operand should be passed as a float value in the args array
+ * right after the coord vector. After packing it's not used anymore,
+ * that's why arg_count is not increased */
+ coords[4] = lp_build_emit_fetch(bld_base, inst, 1, 0);
+ }
+
+ if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
+ inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) &&
+ inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
+ radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
+ }
+
+ for (chan = 0; chan < 4; chan++ ) {
+ coords[chan] = LLVMBuildBitCast(gallivm->builder,
+ coords[chan],
+ LLVMInt32TypeInContext(gallivm->context),
+ "");
+ }
+
+ emit_data->args[1] = lp_build_gather_values(gallivm, coords, 4);
/* Resource */
- emit_data->args[2] = use_sgpr(bld_base->base.gallivm, SGPR_I64, 2);
- emit_data->args[3] = lp_build_const_int32(bld_base->base.gallivm,
- 8 * emit_data->inst->Src[1].Register.Index);
+ ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V8I32, SI_SGPR_RESOURCE);
+ offset = lp_build_const_int32(bld_base->base.gallivm,
+ emit_data->inst->Src[1].Register.Index);
+ emit_data->args[2] = build_indexed_load(bld_base->base.gallivm,
+ ptr, offset);
/* Sampler */
- emit_data->args[4] = use_sgpr(bld_base->base.gallivm, SGPR_I64, 1);
- emit_data->args[5] = lp_build_const_int32(bld_base->base.gallivm,
- 4 * emit_data->inst->Src[1].Register.Index);
+ ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V4I32, SI_SGPR_SAMPLER);
+ offset = lp_build_const_int32(bld_base->base.gallivm,
+ emit_data->inst->Src[1].Register.Index);
+ emit_data->args[3] = build_indexed_load(bld_base->base.gallivm,
+ ptr, offset);
/* Dimensions */
- /* XXX: We might want to pass this information to the shader at some. */
-/* emit_data->args[4] = lp_build_const_int32(bld_base->base.gallivm,
+ emit_data->args[4] = lp_build_const_int32(bld_base->base.gallivm,
emit_data->inst->Texture.Texture);
-*/
- emit_data->arg_count = 6;
+ emit_data->arg_count = 5;
/* XXX: To optimize, we could use a float or v2f32, if the last bits of
* the writemask are clear */
emit_data->dst_type = LLVMVectorType(
static const struct lp_build_tgsi_action tex_action = {
.fetch_args = tex_fetch_args,
.emit = lp_build_tgsi_intrinsic,
- .intr_name = "llvm.SI.sample"
+ .intr_name = "llvm.SI.sample."
+};
+
+static const struct lp_build_tgsi_action txb_action = {
+ .fetch_args = tex_fetch_args,
+ .emit = lp_build_tgsi_intrinsic,
+ .intr_name = "llvm.SI.sampleb."
+};
+
+static const struct lp_build_tgsi_action txl_action = {
+ .fetch_args = tex_fetch_args,
+ .emit = lp_build_tgsi_intrinsic,
+ .intr_name = "llvm.SI.samplel."
};
int si_pipe_shader_create(
struct pipe_context *ctx,
- struct si_pipe_shader *shader)
+ struct si_pipe_shader *shader,
+ struct si_shader_key key)
{
struct r600_context *rctx = (struct r600_context*)ctx;
+ struct si_pipe_shader_selector *sel = shader->selector;
struct si_shader_context si_shader_ctx;
struct tgsi_shader_info shader_info;
struct lp_build_tgsi_context * bld_base;
unsigned char * inst_bytes;
unsigned inst_byte_count;
unsigned i;
+ uint32_t *ptr;
+ bool dump;
+
+ dump = debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE);
+
+ assert(shader->shader.noutput == 0);
+ assert(shader->shader.ninterp == 0);
+ assert(shader->shader.ninput == 0);
+ memset(&si_shader_ctx, 0, sizeof(si_shader_ctx));
radeon_llvm_context_init(&si_shader_ctx.radeon_bld);
bld_base = &si_shader_ctx.radeon_bld.soa.bld_base;
- tgsi_scan_shader(shader->tokens, &shader_info);
+ tgsi_scan_shader(sel->tokens, &shader_info);
+ if (shader_info.indirect_files != 0) {
+ fprintf(stderr, "Indirect addressing not fully handled yet\n");
+ return -ENOSYS;
+ }
+
+ shader->shader.uses_kill = shader_info.uses_kill;
bld_base->info = &shader_info;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
bld_base->emit_prologue = si_llvm_emit_prologue;
bld_base->emit_epilogue = si_llvm_emit_epilogue;
bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
+ bld_base->op_actions[TGSI_OPCODE_TXB] = txb_action;
+ bld_base->op_actions[TGSI_OPCODE_TXL] = txl_action;
+ bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
si_shader_ctx.radeon_bld.load_input = declare_input;
- si_shader_ctx.tokens = shader->tokens;
+ si_shader_ctx.tokens = sel->tokens;
tgsi_parse_init(&si_shader_ctx.parse, si_shader_ctx.tokens);
si_shader_ctx.shader = shader;
+ si_shader_ctx.key = key;
si_shader_ctx.type = si_shader_ctx.parse.FullHeader.Processor.Processor;
si_shader_ctx.rctx = rctx;
- shader->shader.nr_cbufs = rctx->nr_cbufs;
+ shader->shader.nr_cbufs = rctx->framebuffer.nr_cbufs;
+
+ /* Dump TGSI code before doing TGSI->LLVM conversion in case the
+ * conversion fails. */
+ if (dump) {
+ tgsi_dump(sel->tokens, 0);
+ }
- lp_build_tgsi_llvm(bld_base, shader->tokens);
+ if (!lp_build_tgsi_llvm(bld_base, sel->tokens)) {
+ fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n");
+ return -EINVAL;
+ }
radeon_llvm_finalize_module(&si_shader_ctx.radeon_bld);
mod = bld_base->base.gallivm->module;
- tgsi_dump(shader->tokens, 0);
- LLVMDumpModule(mod);
- radeon_llvm_compile(mod, &inst_bytes, &inst_byte_count, "SI", 1 /* dump */);
- fprintf(stderr, "SI CODE:\n");
- for (i = 0; i < inst_byte_count; i+=4 ) {
- fprintf(stderr, "%02x%02x%02x%02x\n", inst_bytes[i + 3],
- inst_bytes[i + 2], inst_bytes[i + 1],
- inst_bytes[i]);
+ if (dump) {
+ LLVMDumpModule(mod);
+ }
+ radeon_llvm_compile(mod, &inst_bytes, &inst_byte_count, "SI", dump);
+ if (dump) {
+ fprintf(stderr, "SI CODE:\n");
+ for (i = 0; i < inst_byte_count; i+=4 ) {
+ fprintf(stderr, "%02x%02x%02x%02x\n", inst_bytes[i + 3],
+ inst_bytes[i + 2], inst_bytes[i + 1],
+ inst_bytes[i]);
+ }
}
shader->num_sgprs = util_le32_to_cpu(*(uint32_t*)inst_bytes);
shader->num_vgprs = util_le32_to_cpu(*(uint32_t*)(inst_bytes + 4));
shader->spi_ps_input_ena = util_le32_to_cpu(*(uint32_t*)(inst_bytes + 8));
+ radeon_llvm_dispose(&si_shader_ctx.radeon_bld);
tgsi_parse_free(&si_shader_ctx.parse);
/* copy new shader */
+ si_resource_reference(&shader->bo, NULL);
+ shader->bo = si_resource_create_custom(ctx->screen, PIPE_USAGE_IMMUTABLE,
+ inst_byte_count - 12);
if (shader->bo == NULL) {
- uint32_t *ptr;
+ return -ENOMEM;
+ }
- shader->bo = (struct r600_resource*)
- pipe_buffer_create(ctx->screen, PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE, inst_byte_count);
- if (shader->bo == NULL) {
- return -ENOMEM;
+ ptr = (uint32_t*)rctx->ws->buffer_map(shader->bo->cs_buf, rctx->cs, PIPE_TRANSFER_WRITE);
+ if (0 /*R600_BIG_ENDIAN*/) {
+ for (i = 0; i < (inst_byte_count-12)/4; ++i) {
+ ptr[i] = util_bswap32(*(uint32_t*)(inst_bytes+12 + i*4));
}
- ptr = (uint32_t*)rctx->ws->buffer_map(shader->bo->cs_buf, rctx->cs, PIPE_TRANSFER_WRITE);
- if (0 /*R600_BIG_ENDIAN*/) {
- for (i = 0; i < (inst_byte_count-12)/4; ++i) {
- ptr[i] = util_bswap32(*(uint32_t*)(inst_bytes+12 + i*4));
- }
- } else {
- memcpy(ptr, inst_bytes + 12, inst_byte_count - 12);
- }
- rctx->ws->buffer_unmap(shader->bo->cs_buf);
+ } else {
+ memcpy(ptr, inst_bytes + 12, inst_byte_count - 12);
}
+ rctx->ws->buffer_unmap(shader->bo->cs_buf);
free(inst_bytes);
void si_pipe_shader_destroy(struct pipe_context *ctx, struct si_pipe_shader *shader)
{
- pipe_resource_reference((struct pipe_resource**)&shader->bo, NULL);
-
- memset(&shader->shader,0,sizeof(struct r600_shader));
+ si_resource_reference(&shader->bo, NULL);
}