[Mesa-dev] [PATCH 3/7] radeon/llvm: make SGPRs proper function arguments
Christian König
deathsimple at vodafone.de
Tue Mar 5 06:27:21 PST 2013
From: Christian König <christian.koenig at amd.com>
Signed-off-by: Christian König <christian.koenig at amd.com>
---
src/gallium/drivers/r600/r600_llvm.c | 1 +
src/gallium/drivers/radeon/radeon_llvm.h | 3 +
.../drivers/radeon/radeon_setup_tgsi_llvm.c | 32 +++---
src/gallium/drivers/radeonsi/radeonsi_shader.c | 106 ++++++--------------
src/gallium/drivers/radeonsi/radeonsi_shader.h | 11 ++
5 files changed, 66 insertions(+), 87 deletions(-)
diff --git a/src/gallium/drivers/r600/r600_llvm.c b/src/gallium/drivers/r600/r600_llvm.c
index aee7cea..7fe15be 100644
--- a/src/gallium/drivers/r600/r600_llvm.c
+++ b/src/gallium/drivers/r600/r600_llvm.c
@@ -524,6 +524,7 @@ LLVMModuleRef r600_tgsi_llvm(
struct tgsi_shader_info shader_info;
struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
radeon_llvm_context_init(ctx);
+ radeon_llvm_create_func(ctx, NULL, 0);
tgsi_scan_shader(tokens, &shader_info);
bld_base->info = &shader_info;
diff --git a/src/gallium/drivers/radeon/radeon_llvm.h b/src/gallium/drivers/radeon/radeon_llvm.h
index b1e025b..4c01d68 100644
--- a/src/gallium/drivers/radeon/radeon_llvm.h
+++ b/src/gallium/drivers/radeon/radeon_llvm.h
@@ -160,6 +160,9 @@ void radeon_llvm_emit_prepare_cube_coords(struct lp_build_tgsi_context * bld_bas
void radeon_llvm_context_init(struct radeon_llvm_context * ctx);
+void radeon_llvm_create_func(struct radeon_llvm_context * ctx,
+ LLVMTypeRef *ParamTypes, unsigned ParamCount);
+
void radeon_llvm_dispose(struct radeon_llvm_context * ctx);
inline static struct radeon_llvm_context * radeon_llvm_context(
diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
index 6103d37..f7b7586 100644
--- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
+++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
@@ -754,6 +754,8 @@ static void emit_icmp(
case TGSI_OPCODE_ISLT: pred = LLVMIntSLT; break;
default:
assert(!"unknown instruction");
+ pred = 0;
+ break;
}
LLVMValueRef v = LLVMBuildICmp(builder, pred,
@@ -770,10 +772,7 @@ static void emit_ucmp(
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
- unsigned pred;
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- LLVMContextRef context = bld_base->base.gallivm->context;
-
LLVMValueRef v = LLVMBuildFCmp(builder, LLVMRealUGE,
emit_data->args[0], lp_build_const_float(bld_base->base.gallivm, 0.), "");
@@ -802,7 +801,7 @@ static void emit_cmp(
case TGSI_OPCODE_SLT: pred = LLVMRealULT; break;
case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
case TGSI_OPCODE_SGT: pred = LLVMRealUGT; break;
- default: assert(!"unknown instruction");
+ default: assert(!"unknown instruction"); pred = 0; break;
}
cond = LLVMBuildFCmp(builder,
@@ -1095,8 +1094,6 @@ static void build_tgsi_intrinsic_readonly(
void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
{
struct lp_type type;
- LLVMTypeRef main_fn_type;
- LLVMBasicBlockRef main_fn_body;
/* Initialize the gallivm object:
* We are only using the module, context, and builder fields of this struct.
@@ -1110,14 +1107,6 @@ void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
ctx->gallivm.context);
ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
- /* Setup the module */
- main_fn_type = LLVMFunctionType(LLVMVoidTypeInContext(ctx->gallivm.context),
- NULL, 0, 0);
- ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
- main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
- ctx->main_fn, "main_body");
- LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
-
ctx->store_output_intr = "llvm.AMDGPU.store.output.";
ctx->swizzle_intr = "llvm.AMDGPU.swizzle";
struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
@@ -1262,6 +1251,21 @@ void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
bld_base->rsq_action.intr_name = "llvm.AMDGPU.rsq";
}
+void radeon_llvm_create_func(struct radeon_llvm_context * ctx,
+ LLVMTypeRef *ParamTypes, unsigned ParamCount)
+{
+ LLVMTypeRef main_fn_type;
+ LLVMBasicBlockRef main_fn_body;
+
+ /* Setup the function */
+ main_fn_type = LLVMFunctionType(LLVMVoidTypeInContext(ctx->gallivm.context),
+ ParamTypes, ParamCount, 0);
+ ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
+ main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
+ ctx->main_fn, "main_body");
+ LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
+}
+
void radeon_llvm_finalize_module(struct radeon_llvm_context * ctx)
{
struct gallivm_state * gallivm = ctx->soa.bld_base.base.gallivm;
diff --git a/src/gallium/drivers/radeonsi/radeonsi_shader.c b/src/gallium/drivers/radeonsi/radeonsi_shader.c
index af05078..9e1e25e 100644
--- a/src/gallium/drivers/radeonsi/radeonsi_shader.c
+++ b/src/gallium/drivers/radeonsi/radeonsi_shader.c
@@ -82,13 +82,6 @@ static struct si_shader_context * si_shader_context(
#define CONST_ADDR_SPACE 2
#define USER_SGPR_ADDR_SPACE 8
-enum sgpr_type {
- SGPR_CONST_PTR_F32,
- SGPR_CONST_PTR_V16I8,
- SGPR_CONST_PTR_V32I8,
- SGPR_I32
-};
-
/**
* Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad
*
@@ -112,66 +105,6 @@ static LLVMValueRef build_indexed_load(
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;
- LLVMTypeRef ret_type;
- LLVMValueRef ptr;
-
- sgpr_index = lp_build_const_int32(gallivm, sgpr);
-
- 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;
-
- case SGPR_I32:
- ret_type = LLVMInt32TypeInContext(gallivm->context);
- break;
-
- case SGPR_CONST_PTR_V16I8:
- assert(sgpr % 2 == 0);
- ret_type = LLVMInt8TypeInContext(gallivm->context);
- ret_type = LLVMVectorType(ret_type, 16);
- ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
- break;
-
- case SGPR_CONST_PTR_V32I8:
- assert(sgpr % 2 == 0);
- ret_type = LLVMInt8TypeInContext(gallivm->context);
- ret_type = LLVMVectorType(ret_type, 32);
- 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(
struct si_shader_context * si_shader_ctx,
unsigned input_index,
@@ -191,7 +124,7 @@ static void declare_input_vs(
unsigned chan;
/* Load the T list */
- t_list_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_V16I8, SI_SGPR_VERTEX_BUFFER);
+ t_list_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_VERTEX_BUFFER);
t_offset = lp_build_const_int32(base->gallivm, input_index);
@@ -244,7 +177,7 @@ static void declare_input_fs(
* [32:16] ParamOffset
*
*/
- LLVMValueRef params = use_sgpr(base->gallivm, SGPR_I32, SI_PS_NUM_USER_SGPR);
+ LLVMValueRef params = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_PRIM_MASK);
LLVMValueRef attr_number;
if (decl->Semantic.Name == TGSI_SEMANTIC_POSITION) {
@@ -419,6 +352,7 @@ static LLVMValueRef fetch_constant(
enum tgsi_opcode_type type,
unsigned swizzle)
{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
struct lp_build_context * base = &bld_base->base;
unsigned idx;
@@ -442,7 +376,7 @@ static LLVMValueRef fetch_constant(
return bitcast(bld_base, type, load);
}
- const_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_F32, SI_SGPR_CONST);
+ const_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_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
@@ -804,6 +738,7 @@ static void tex_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;
unsigned opcode = inst->Instruction.Opcode;
@@ -927,14 +862,14 @@ static void tex_fetch_args(
emit_data->args[1] = lp_build_gather_values(gallivm, address, count);
/* Resource */
- ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V32I8, SI_SGPR_RESOURCE);
+ ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_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 */
- ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V16I8, SI_SGPR_SAMPLER);
+ ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_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,
@@ -984,6 +919,31 @@ static const struct lp_build_tgsi_action txl_action = {
.intr_name = "llvm.SI.samplel."
};
+static void create_function(struct si_shader_context *si_shader_ctx)
+{
+ struct gallivm_state *gallivm = si_shader_ctx->radeon_bld.soa.bld_base.base.gallivm;
+ LLVMTypeRef params[4], f, i8;
+ unsigned i;
+
+ f = LLVMFloatTypeInContext(gallivm->context);
+ i8 = LLVMInt8TypeInContext(gallivm->context);
+ params[SI_PARAM_CONST] = LLVMPointerType(f, CONST_ADDR_SPACE);
+ params[SI_PARAM_SAMPLER] = LLVMPointerType(LLVMVectorType(i8, 16), CONST_ADDR_SPACE);
+ params[SI_PARAM_RESOURCE] = LLVMPointerType(LLVMVectorType(i8, 32), CONST_ADDR_SPACE);
+
+ if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX)
+ params[SI_PARAM_VERTEX_BUFFER] = params[SI_PARAM_SAMPLER];
+ else
+ params[SI_PARAM_PRIM_MASK] = LLVMInt32TypeInContext(gallivm->context);
+
+ radeon_llvm_create_func(&si_shader_ctx->radeon_bld, params, 4);
+
+ radeon_llvm_shader_type(si_shader_ctx->radeon_bld.main_fn, si_shader_ctx->type);
+ for (i = SI_PARAM_CONST; i <= SI_PARAM_VERTEX_BUFFER; ++i) {
+ LLVMValueRef P = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, i);
+ LLVMAddAttribute(P, LLVMInRegAttribute);
+ }
+}
int si_pipe_shader_create(
struct pipe_context *ctx,
@@ -1036,7 +996,7 @@ int si_pipe_shader_create(
si_shader_ctx.type = si_shader_ctx.parse.FullHeader.Processor.Processor;
si_shader_ctx.rctx = rctx;
- radeon_llvm_shader_type(si_shader_ctx.radeon_bld.main_fn, si_shader_ctx.type);
+ create_function(&si_shader_ctx);
shader->shader.nr_cbufs = rctx->framebuffer.nr_cbufs;
diff --git a/src/gallium/drivers/radeonsi/radeonsi_shader.h b/src/gallium/drivers/radeonsi/radeonsi_shader.h
index f54f67c..3712347 100644
--- a/src/gallium/drivers/radeonsi/radeonsi_shader.h
+++ b/src/gallium/drivers/radeonsi/radeonsi_shader.h
@@ -37,6 +37,17 @@
#define SI_VS_NUM_USER_SGPR 8
#define SI_PS_NUM_USER_SGPR 6
+/* LLVM function parameter indices */
+#define SI_PARAM_CONST 0
+#define SI_PARAM_SAMPLER 1
+#define SI_PARAM_RESOURCE 2
+
+/* VS only parameters */
+#define SI_PARAM_VERTEX_BUFFER 3
+
+/* PS only parameters */
+#define SI_PARAM_PRIM_MASK 3
+
struct si_shader_io {
unsigned name;
int sid;
--
1.7.9.5
More information about the mesa-dev
mailing list