KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ac: unify primitive export code 

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
This commit is contained in:
Marek Olšák 2019-12-23 19:26:46 -05:00
parent 1c77a18cc2
commit d1c8aeb24f
4 changed files with 78 additions and 113 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

53
src/amd/llvm/ac_llvm_build.c
View File

@ -4748,6 +4748,59 @@ void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wav
ac_build_endif(ctx, 5020);
}
LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
const struct ac_ngg_prim *prim)
{
/* The prim export format is:
* - bits 0..8: index 0
* - bit 9: edge flag 0
* - bits 10..18: index 1
* - bit 19: edge flag 1
* - bits 20..28: index 2
* - bit 29: edge flag 2
* - bit 31: null primitive (skip)
*/
LLVMBuilderRef builder = ctx->builder;
LLVMValueRef tmp = LLVMBuildZExt(builder, prim->isnull, ctx->i32, "");
LLVMValueRef result = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->i32, 31, false), "");
for (unsigned i = 0; i < prim->num_vertices; ++i) {
tmp = LLVMBuildShl(builder, prim->index[i],
LLVMConstInt(ctx->i32, 10 * i, false), "");
result = LLVMBuildOr(builder, result, tmp, "");
tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->i32, "");
tmp = LLVMBuildShl(builder, tmp,
LLVMConstInt(ctx->i32, 10 * i + 9, false), "");
result = LLVMBuildOr(builder, result, tmp, "");
}
return result;
}
void ac_build_export_prim(struct ac_llvm_context *ctx,
const struct ac_ngg_prim *prim)
{
struct ac_export_args args;
if (prim->passthrough) {
args.out[0] = prim->passthrough;
} else {
args.out[0] = ac_pack_prim_export(ctx, prim);
}
args.out[0] = LLVMBuildBitCast(ctx->builder, args.out[0], ctx->f32, "");
args.out[1] = LLVMGetUndef(ctx->f32);
args.out[2] = LLVMGetUndef(ctx->f32);
args.out[3] = LLVMGetUndef(ctx->f32);
args.target = V_008DFC_SQ_EXP_PRIM;
args.enabled_channels = 1;
args.done = true;
args.valid_mask = false;
args.compr = false;
ac_build_export(ctx, &args);
}
static LLVMTypeRef
arg_llvm_type(enum ac_arg_type type, unsigned size, struct ac_llvm_context *ctx)
{

13
src/amd/llvm/ac_llvm_build.h
View File

@ -750,6 +750,19 @@ ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt);
struct ac_ngg_prim {
unsigned num_vertices;
LLVMValueRef isnull;
LLVMValueRef index[3];
LLVMValueRef edgeflag[3];
LLVMValueRef passthrough;
};
LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
const struct ac_ngg_prim *prim);
void ac_build_export_prim(struct ac_llvm_context *ctx,
const struct ac_ngg_prim *prim);
static inline LLVMValueRef
ac_get_arg(struct ac_llvm_context *ctx, struct ac_arg arg)
{

60
src/amd/vulkan/radv_nir_to_llvm.c
View File

@ -2371,47 +2371,6 @@ ngg_gs_emit_vertex_ptr(struct radv_shader_context *ctx, LLVMValueRef gsthread,
return ngg_gs_vertex_ptr(ctx, vertexidx);
}
struct ngg_prim {
unsigned num_vertices;
LLVMValueRef isnull;
LLVMValueRef index[3];
LLVMValueRef edgeflag[3];
};
static void build_export_prim(struct radv_shader_context *ctx,
const struct ngg_prim *prim)
{
LLVMBuilderRef builder = ctx->ac.builder;
struct ac_export_args args;
LLVMValueRef tmp;
tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");
for (unsigned i = 0; i < prim->num_vertices; ++i) {
tmp = LLVMBuildShl(builder, prim->index[i],
LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
tmp = LLVMBuildShl(builder, tmp,
LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
}
args.out[0] = LLVMBuildBitCast(builder, args.out[0], ctx->ac.f32, "");
args.out[1] = LLVMGetUndef(ctx->ac.f32);
args.out[2] = LLVMGetUndef(ctx->ac.f32);
args.out[3] = LLVMGetUndef(ctx->ac.f32);
args.target = V_008DFC_SQ_EXP_PRIM;
args.enabled_channels = 1;
args.done = true;
args.valid_mask = false;
args.compr = false;
ac_build_export(&ctx->ac, &args);
}
static struct radv_stream_output *
radv_get_stream_output_by_loc(struct radv_streamout_info *so, unsigned location)
{
@ -3002,14 +2961,7 @@ handle_ngg_outputs_post_2(struct radv_shader_context *ctx)
ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
/* TODO: streamout queries */
/* Export primitive data to the index buffer. Format is:
* - bits 0..8: index 0
* - bit 9: edge flag 0
* - bits 10..18: index 1
* - bit 19: edge flag 1
* - bits 20..28: index 2
* - bit 29: edge flag 2
* - bit 31: null primitive (skip)
/* Export primitive data to the index buffer.
*
* For the first version, we will always build up all three indices
* independent of the primitive type. The additional garbage data
@ -3020,7 +2972,7 @@ handle_ngg_outputs_post_2(struct radv_shader_context *ctx)
*/
ac_build_ifcc(&ctx->ac, is_gs_thread, 6001);
{
struct ngg_prim prim = {};
struct ac_ngg_prim prim = {};
prim.num_vertices = num_vertices;
prim.isnull = ctx->ac.i1false;
@ -3033,7 +2985,7 @@ handle_ngg_outputs_post_2(struct radv_shader_context *ctx)
prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
}
build_export_prim(ctx, &prim);
ac_build_export_prim(&ctx->ac, &prim);
}
ac_build_endif(&ctx->ac, 6001);
@ -3323,7 +3275,7 @@ static void gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context *ctx)
ac_build_ifcc(&ctx->ac, tmp, 5140);
{
LLVMValueRef flags;
struct ngg_prim prim = {};
struct ac_ngg_prim prim = {};
prim.num_vertices = verts_per_prim;
tmp = ngg_gs_vertex_ptr(ctx, tid);
@ -3352,7 +3304,7 @@ static void gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context *ctx)
LLVMValueRef is_odd = LLVMBuildLShr(builder, flags, ctx->ac.i8_1, "");
is_odd = LLVMBuildTrunc(builder, is_odd, ctx->ac.i1, "");
struct ngg_prim in = prim;
struct ac_ngg_prim in = prim;
prim.index[0] = in.index[0];
prim.index[1] = LLVMBuildSelect(builder, is_odd,
in.index[2], in.index[1], "");
@ -3360,7 +3312,7 @@ static void gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context *ctx)
in.index[1], in.index[2], "");
}
build_export_prim(ctx, &prim);
ac_build_export_prim(&ctx->ac, &prim);
}
ac_build_endif(&ctx->ac, 5140);

65
src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
View File

@ -71,52 +71,6 @@ static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
LLVMConstInt(ctx->i32, GFX10_GS_QUERY_BUF, false));
}
struct ngg_prim {
unsigned num_vertices;
LLVMValueRef isnull;
LLVMValueRef index[3];
LLVMValueRef edgeflag[3];
LLVMValueRef passthrough;
};
static void build_export_prim(struct si_shader_context *ctx,
const struct ngg_prim *prim)
{
LLVMBuilderRef builder = ctx->ac.builder;
struct ac_export_args args;
LLVMValueRef tmp;
if (prim->passthrough) {
args.out[0] = prim->passthrough;
} else {
tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");
for (unsigned i = 0; i < prim->num_vertices; ++i) {
tmp = LLVMBuildShl(builder, prim->index[i],
LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
tmp = LLVMBuildShl(builder, tmp,
LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
}
}
args.out[0] = LLVMBuildBitCast(builder, args.out[0], ctx->ac.f32, "");
args.out[1] = LLVMGetUndef(ctx->ac.f32);
args.out[2] = LLVMGetUndef(ctx->ac.f32);
args.out[3] = LLVMGetUndef(ctx->ac.f32);
args.target = V_008DFC_SQ_EXP_PRIM;
args.enabled_channels = 1;
args.done = true;
args.valid_mask = false;
args.compr = false;
ac_build_export(&ctx->ac, &args);
}
static void build_streamout_vertex(struct si_shader_context *ctx,
LLVMValueRef *so_buffer, LLVMValueRef *wg_offset_dw,
unsigned stream, LLVMValueRef offset_vtx,
@ -693,14 +647,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
ac_build_endif(&ctx->ac, 5029);
}
/* Export primitive data to the index buffer. Format is:
* - bits 0..8: index 0
* - bit 9: edge flag 0
* - bits 10..18: index 1
* - bit 19: edge flag 1
* - bits 20..28: index 2
* - bit 29: edge flag 2
* - bit 31: null primitive (skip)
/* Build the primitive export.
*
* For the first version, we will always build up all three indices
* independent of the primitive type. The additional garbage data
@ -711,7 +658,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
*/
ac_build_ifcc(&ctx->ac, is_gs_thread, 6001);
{
struct ngg_prim prim = {};
struct ac_ngg_prim prim = {};
if (gfx10_is_ngg_passthrough(ctx->shader)) {
prim.passthrough = ac_get_arg(&ctx->ac, ctx->gs_vtx01_offset);
@ -739,7 +686,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
}
}
build_export_prim(ctx, &prim);
ac_build_export_prim(&ctx->ac, &prim);
}
ac_build_endif(&ctx->ac, 6001);
@ -1213,7 +1160,7 @@ void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
ac_build_ifcc(&ctx->ac, tmp, 5140);
{
LLVMValueRef flags;
struct ngg_prim prim = {};
struct ac_ngg_prim prim = {};
prim.num_vertices = verts_per_prim;
tmp = ngg_gs_vertex_ptr(ctx, tid);
@ -1242,7 +1189,7 @@ void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
si_unpack_param(ctx, ctx->vs_state_bits, 4, 2),
ctx->i32_0, "");
struct ngg_prim in = prim;
struct ac_ngg_prim in = prim;
prim.index[0] = LLVMBuildSelect(builder, flatshade_first,
in.index[0],
LLVMBuildSelect(builder, is_odd,
@ -1258,7 +1205,7 @@ void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
in.index[2], "");
}
build_export_prim(ctx, &prim);
ac_build_export_prim(&ctx->ac, &prim);
}
ac_build_endif(&ctx->ac, 5140);