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

radeonsi: use LLVMBuildLoad2 for inter-stage outputs loads 

The PS case was covered by the previous commit, so we can use f32
everywhere.

Reviewed-by: Mihai Preda <mhpreda@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17361>
This commit is contained in:
Pierre-Eric Pelloux-Prayer 2022-07-05 13:44:24 +02:00 committed by Marge Bot
parent 248781dea1
commit 01314d0880
4 changed files with 24 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
View File

@ -1044,7 +1044,7 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
pos_index = i;
for (unsigned j = 0; j < 4; j++) {
position[j] = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + j], "");
position[j] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addrs[4 * i + j], "");
}
/* Store Position.W into LDS. */
@ -1071,7 +1071,7 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
if (!(clipdist_enable & BITFIELD_BIT(index)))
continue;
LLVMValueRef distance = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + j], "");
LLVMValueRef distance = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addrs[4 * i + j], "");
add_clipdist_bit(ctx, distance, index, &packed_data);
has_clipdist_mask = true;
}
@ -1079,7 +1079,7 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
case VARYING_SLOT_CLIP_VERTEX:
for (unsigned j = 0; j < 4; j++)
clipvertex[j] = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + j], "");
clipvertex[j] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addrs[4 * i + j], "");
if (add_clipdist_bits_for_clipvertex(ctx, clipdist_enable, clipvertex, &packed_data))
has_clipdist_mask = true;
@ -1291,7 +1291,8 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
/* Store Position.XYZW into LDS. */
for (unsigned chan = 0; chan < 4; chan++) {
LLVMBuildStore(
builder, ac_to_integer(&ctx->ac, LLVMBuildLoad(builder, addrs[4 * pos_index + chan], "")),
builder, ac_to_integer(&ctx->ac,
LLVMBuildLoad2(builder, ctx->ac.f32, addrs[4 * pos_index + chan], "")),
ac_build_gep0(&ctx->ac, new_vtx, LLVMConstInt(ctx->ac.i32, lds_pos_x + chan, 0)));
}
@ -1532,15 +1533,16 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
*/
if (ctx->so.num_outputs) {
tmp = ac_build_gep0(&ctx->ac, vertex_ptr, LLVMConstInt(ctx->ac.i32, 4 * i + j, false));
tmp2 = LLVMBuildLoad(builder, addrs[4 * i + j], "");
tmp2 = ac_to_integer(&ctx->ac, tmp2);
tmp2 = LLVMBuildLoad2(builder, ctx->ac.f32, addrs[4 * i + j], "");
LLVMTypeRef type = ac_to_integer_type(&ctx->ac, ctx->ac.f32);
tmp2 = LLVMBuildBitCast(ctx->ac.builder, tmp2, type, "");
LLVMBuildStore(builder, tmp2, tmp);
}
}
/* Store the edgeflag at the end (if streamout is enabled) */
if (info->output_semantic[i] == VARYING_SLOT_EDGE && gfx10_ngg_writes_user_edgeflags(ctx->shader)) {
LLVMValueRef edgeflag = LLVMBuildLoad(builder, addrs[4 * i], "");
LLVMValueRef edgeflag = LLVMBuildLoad2(builder, ctx->ac.f32, addrs[4 * i], "");
/* The output is a float, but the hw expects a 1-bit integer. */
edgeflag = LLVMBuildFPToUI(ctx->ac.builder, edgeflag, ctx->ac.i32, "");
edgeflag = ac_build_umin(&ctx->ac, edgeflag, ctx->ac.i32_1);
@ -1706,11 +1708,12 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
tmp = LLVMConstInt(ctx->ac.i32, lds_pos_x + j, 0);
tmp = ac_build_gep0(&ctx->ac, vertex_ptr, tmp);
tmp = LLVMBuildLoad2(builder, ctx->ac.i32, tmp, "");
outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
outputs[i].values[j] = LLVMBuildBitCast(ctx->ac.builder, tmp,
ac_to_float_type(&ctx->ac, ctx->ac.i32), "");
}
} else {
for (unsigned j = 0; j < 4; j++) {
outputs[i].values[j] = LLVMBuildLoad(builder, addrs[4 * i + j], "");
outputs[i].values[j] = LLVMBuildLoad2(builder, ctx->ac.f32, addrs[4 * i + j], "");
}
}
}
@ -1732,7 +1735,7 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
outputs[i].values[0] = si_get_primitive_id(ctx, 0);
}
outputs[i].values[0] = ac_to_float(&ctx->ac, outputs[i].values[0]);
outputs[i].values[0] = LLVMBuildBitCast(ctx->ac.builder, outputs[i].values[0], ctx->ac.f32, "");
for (unsigned j = 1; j < 4; j++)
outputs[i].values[j] = LLVMGetUndef(ctx->ac.f32);
i++;
@ -1870,8 +1873,9 @@ void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx, unsigned stream, LL
((info->output_streams[i] >> (2 * chan)) & 3) != stream)
continue;
LLVMValueRef out_val = LLVMBuildLoad(builder, addrs[4 * i + chan], "");
out_val = ac_to_integer(&ctx->ac, out_val);
LLVMValueRef out_val = LLVMBuildLoad2(builder, ctx->ac.f32, addrs[4 * i + chan], "");
LLVMTypeRef as_int = ac_to_integer_type(&ctx->ac, ctx->ac.f32);
out_val = LLVMBuildBitCast(ctx->ac.builder, out_val, as_int, "");
LLVMBuildStore(builder, out_val, ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx));
}
}
@ -2304,7 +2308,8 @@ void gfx10_ngg_gs_build_end(struct si_shader_context *ctx)
for (unsigned j = 0; j < 4; j++, out_idx++) {
tmp = ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx);
tmp = LLVMBuildLoad(builder, tmp, "");
tmp = LLVMBuildLoad2(builder, ctx->ac.i32, tmp, "");
assert(LLVMGetTypeKind(LLVMTypeOf(tmp)) != LLVMPointerTypeKind);
outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
outputs[i].vertex_streams = info->output_streams[i];
}

2
src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
View File

@ -210,7 +210,7 @@ static void si_llvm_emit_vertex(struct ac_shader_abi *abi, unsigned stream, LLVM
((info->output_streams[i] >> (2 * chan)) & 3) != stream)
continue;
LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
LLVMValueRef out_val = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addrs[4 * i + chan], "");
LLVMValueRef voffset =
LLVMConstInt(ctx->ac.i32, offset * shader->selector->info.base.gs.vertices_out, 0);
offset++;

4
src/gallium/drivers/radeonsi/si_shader_llvm_tess.c
View File

@ -523,7 +523,7 @@ void si_llvm_tcs_build_end(struct si_shader_context *ctx)
for (unsigned i = 0; i < 6; i++) {
int loc = i < 4 ? outer_loc : inner_loc;
LLVMValueRef value = loc < 0 ? LLVMGetUndef(ctx->ac.f32) :
LLVMBuildLoad(builder, ctx->abi.outputs[loc * 4 + i % 4], "");
LLVMBuildLoad2(builder, ctx->ac.f32, ctx->abi.outputs[loc * 4 + i % 4], "");
value = ac_to_float(&ctx->ac, value);
ret = LLVMBuildInsertValue(builder, ret, value, vgpr++, "");
}
@ -585,7 +585,7 @@ void si_llvm_ls_build_end(struct si_shader_context *ctx)
if (!(info->output_usagemask[i] & (1 << chan)))
continue;
LLVMValueRef value = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
LLVMValueRef value = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addrs[4 * i + chan], "");
ctx->return_value = LLVMBuildInsertValue(ctx->ac.builder, ctx->return_value,
value, ret_offset + param * 4 + chan, "");

6
src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
View File

@ -499,7 +499,7 @@ static void si_vertex_color_clamping(struct si_shader_context *ctx,
continue;
for (unsigned j = 0; j < 4; j++) {
outputs[i].values[j] = LLVMBuildLoad(ctx->ac.builder, addr[i][j], "");
outputs[i].values[j] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addr[i][j], "");
}
}
}
@ -794,7 +794,7 @@ void si_llvm_build_vs_exports(struct si_shader_context *ctx, LLVMValueRef num_ex
for (unsigned i = 0; i < shader->info.nr_param_exports; i++) {
for (unsigned chan = 0; chan < 4; chan++) {
param_exports[i].out[chan] =
LLVMBuildLoad(ctx->ac.builder, param_exports[i].out[chan], "");
LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, param_exports[i].out[chan], "");
}
LLVMValueRef vdata = ac_build_gather_values_extended(&ctx->ac, param_exports[i].out,
@ -826,7 +826,7 @@ void si_llvm_vs_build_end(struct si_shader_context *ctx)
outputs[i].semantic = info->output_semantic[i];
for (j = 0; j < 4; j++) {
outputs[i].values[j] = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + j], "");
outputs[i].values[j] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addrs[4 * i + j], "");
outputs[i].vertex_streams = info->output_streams[i];
}
}