#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_tgsi.h"
#include "radeon_llvm.h"
#include "radeonsi_pipe.h"
#include "radeonsi_shader.h"
+#include "si_state.h"
#include "sid.h"
#include <assert.h>
#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;
+ /* Load the T list */
/* 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);
+ t_list_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_V4I32, 6);
- 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 */
"llvm.SI.vs.load.buffer.index", uint->elem_type, NULL, 0);
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++) {
/* 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->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";
LLVMValueRef const_ptr;
LLVMValueRef offset;
+ LLVMValueRef 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, 0);
/* 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 = lp_build_const_int32(base->gallivm,
(reg->Register.Index * 4) + swizzle);
- return lp_build_intrinsic_binary(base->gallivm->builder,
- "llvm.SI.load.const", base->elem_type, const_ptr, offset);
-}
-
-
-/* Declare some intrinsics with the correct attributes */
-static void si_llvm_emit_prologue(struct lp_build_tgsi_context * bld_base)
-{
- LLVMValueRef function;
- struct gallivm_state * gallivm = bld_base->base.gallivm;
-
- LLVMTypeRef i64 = LLVMInt64TypeInContext(gallivm->context);
- LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
-
- /* 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);
-
- /* 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);
+ load = build_indexed_load(base->gallivm, const_ptr, offset);
+ return bitcast(bld_base, type, load);
}
/* 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 last_args[9] = { 0 };
+ unsigned color_count = 0;
+ unsigned param_count = 0;
while (!tgsi_parse_end_of_tokens(parse)) {
/* XXX: component_bits controls which components of the output
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++ ) {
case TGSI_SEMANTIC_COLOR:
if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
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;
break;
case TGSI_SEMANTIC_GENERIC:
target = V_008DFC_SQ_EXP_PARAM + param_count;
+ shader->output[i].param_offset = param_count;
param_count++;
break;
default:
}
}
+ 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;
+ }
+
/* 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)
{
+ const struct tgsi_full_instruction * inst = emit_data->inst;
+ LLVMValueRef ptr;
+ LLVMValueRef offset;
+
/* 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) {
+ LLVMValueRef src_w;
+ unsigned chan;
+ LLVMValueRef coords[4];
+
+ emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
+ src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
+
+ for (chan = 0; chan < 3; chan++ ) {
+ LLVMValueRef arg = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 0, chan);
+ coords[chan] = lp_build_emit_llvm_binary(bld_base,
+ TGSI_OPCODE_DIV,
+ arg, src_w);
+ }
+ coords[3] = bld_base->base.one;
+ emit_data->args[1] = lp_build_gather_values(bld_base->base.gallivm,
+ coords, 4);
+ } else
+ emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
+ 0, LP_CHAN_ALL);
/* 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, 4);
+ 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, 2);
+ 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->inst->Texture.Texture);
*/
- emit_data->arg_count = 6;
+ emit_data->arg_count = 4;
/* XXX: To optimize, we could use a float or v2f32, if the last bits of
* the writemask are clear */
emit_data->dst_type = LLVMVectorType(
unsigned char * inst_bytes;
unsigned inst_byte_count;
unsigned i;
+ uint32_t *ptr;
+ bool dump;
+
+ dump = debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE);
+ memset(&si_shader_ctx.radeon_bld, 0, sizeof(si_shader_ctx.radeon_bld));
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);
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_TXP] = tex_action;
si_shader_ctx.radeon_bld.load_input = declare_input;
si_shader_ctx.tokens = shader->tokens;
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;
- lp_build_tgsi_llvm(bld_base, shader->tokens);
+ /* Dump TGSI code before doing TGSI->LLVM conversion in case the
+ * conversion fails. */
+ if (dump) {
+ tgsi_dump(shader->tokens, 0);
+ }
+
+ if (!lp_build_tgsi_llvm(bld_base, shader->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);
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);
+ si_resource_reference(&shader->bo, NULL);
- memset(&shader->shader,0,sizeof(struct r600_shader));
+ memset(&shader->shader,0,sizeof(struct si_shader));
}