radeonsi: replace llvm legacy gs code with nir lowering

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19489>

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 142e094..adc09f4 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1767,8 +1767,11 @@ static void si_assign_param_offsets(nir_shader *nir, struct si_shader *shader)
    si_nir_assign_param_offsets(nir, shader, slot_remap);
 }
 
-struct nir_shader *si_get_nir_shader(struct si_shader *shader, struct si_shader_args *args,
-                                     bool *free_nir, uint64_t tcs_vgpr_only_inputs)
+struct nir_shader *si_get_nir_shader(struct si_shader *shader,
+                                     struct si_shader_args *args,
+                                     bool *free_nir,
+                                     uint64_t tcs_vgpr_only_inputs,
+                                     ac_nir_gs_output_info *output_info)
 {
    struct si_shader_selector *sel = shader->selector;
    const union si_shader_key *key = &shader->key;
@@ -1916,6 +1919,9 @@ struct nir_shader *si_get_nir_shader(struct si_shader *shader, struct si_shader_
       }
    }
 
+   if (sel->stage == MESA_SHADER_GEOMETRY && !key->ge.as_ngg)
+      NIR_PASS_V(nir, ac_nir_lower_legacy_gs, false, sel->screen->use_ngg, output_info);
+
    NIR_PASS(progress2, nir, si_nir_lower_abi, shader, args);
 
    if (progress2 || opt_offsets)
@@ -2111,7 +2117,8 @@ bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compi
    si_init_shader_args(shader, &args);
 
    bool free_nir;
-   struct nir_shader *nir = si_get_nir_shader(shader, &args, &free_nir, 0);
+   struct nir_shader *nir =
+      si_get_nir_shader(shader, &args, &free_nir, 0, &legacy_gs_output_info.info);
 
    /* Dump NIR before doing NIR->LLVM conversion in case the
     * conversion fails. */

diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 4f7a6b8..3207c52 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -147,8 +147,6 @@ struct si_shader_context {
    LLVMValueRef gs_ngg_emit;
    struct ac_llvm_pointer gs_ngg_scratch;
    LLVMValueRef return_value;
-
-   LLVMValueRef gs_emitted_vertices;
 };
 
 static inline struct si_shader_context *si_shader_context_from_abi(struct ac_shader_abi *abi)
@@ -156,6 +154,9 @@ static inline struct si_shader_context *si_shader_context_from_abi(struct ac_sha
    return container_of(abi, struct si_shader_context, abi);
 }
 
+struct ac_nir_gs_output_info;
+typedef struct ac_nir_gs_output_info ac_nir_gs_output_info;
+
 /* si_shader.c */
 bool si_is_multi_part_shader(struct si_shader *shader);
 bool si_is_merged_shader(struct si_shader *shader);
@@ -169,7 +170,8 @@ void si_get_vs_prolog_key(const struct si_shader_info *info, unsigned num_input_
                           const struct si_vs_prolog_bits *prolog_key,
                           struct si_shader *shader_out, union si_shader_part_key *key);
 struct nir_shader *si_get_nir_shader(struct si_shader *shader, struct si_shader_args *args,
-                                     bool *free_nir, uint64_t tcs_vgpr_only_inputs);
+                                     bool *free_nir, uint64_t tcs_vgpr_only_inputs,
+                                     ac_nir_gs_output_info *output_info);
 void si_get_tcs_epilog_key(struct si_shader *shader, union si_shader_part_key *key);
 bool si_need_ps_prolog(const union si_shader_part_key *key);
 void si_get_ps_prolog_key(struct si_shader *shader, union si_shader_part_key *key,

diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm.c b/src/gallium/drivers/radeonsi/si_shader_llvm.c
index 37f491e..aa9cf56 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader_llvm.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm.c
@@ -849,8 +849,6 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
          LLVMSetAlignment(ctx->gs_ngg_emit, 4);
       } else {
          si_preload_gs_rings(ctx);
-
-         ctx->gs_emitted_vertices = LLVMConstInt(ctx->ac.i32, 0, false);
       }
       break;
 
@@ -1172,7 +1170,8 @@ bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *
          si_init_shader_args(&shader_ls, ctx.args);
 
          bool free_nir;
-         nir = si_get_nir_shader(&shader_ls, ctx.args, &free_nir, sel->info.tcs_vgpr_only_inputs);
+         nir = si_get_nir_shader(&shader_ls, ctx.args, &free_nir,
+                                 sel->info.tcs_vgpr_only_inputs, NULL);
          si_update_shader_binary_info(shader, nir);
 
          if (!si_llvm_translate_nir(&ctx, &shader_ls, nir, free_nir)) {
@@ -1246,7 +1245,7 @@ bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *
          si_init_shader_args(&shader_es, ctx.args);
 
          bool free_nir;
-         nir = si_get_nir_shader(&shader_es, ctx.args, &free_nir, 0);
+         nir = si_get_nir_shader(&shader_es, ctx.args, &free_nir, 0, NULL);
          si_update_shader_binary_info(shader, nir);
 
          if (!si_llvm_translate_nir(&ctx, &shader_es, nir, free_nir)) {

diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
index 36c07d2..30befe5 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
@@ -105,57 +105,9 @@ static LLVMValueRef ngg_get_emulated_counters_buf(struct si_shader_context *ctx)
 
 void si_llvm_gs_build_end(struct si_shader_context *ctx)
 {
-   struct si_shader_info UNUSED *info = &ctx->shader->selector->info;
-
-   assert(info->num_outputs <= AC_LLVM_MAX_OUTPUTS);
-
    if (ctx->screen->info.gfx_level >= GFX10)
       ac_build_waitcnt(&ctx->ac, AC_WAIT_VSTORE);
 
-   if (ctx->screen->use_ngg) {
-      /* Implement PIPE_STAT_QUERY_GS_PRIMITIVES for non-ngg draws because we can't
-       * use pipeline statistics (they would be correct but when screen->use_ngg, we
-       * can't know when the query is started if the next draw(s) will use ngg or not).
-       */
-      LLVMValueRef tmp = GET_FIELD(ctx, GS_STATE_PIPELINE_STATS_EMU);
-      tmp = LLVMBuildTrunc(ctx->ac.builder, tmp, ctx->ac.i1, "");
-      ac_build_ifcc(&ctx->ac, tmp, 5229); /* if (GS_PIPELINE_STATS_EMU) */
-      {
-         LLVMValueRef prim = ctx->ac.i32_0;
-         switch (ctx->shader->selector->info.base.gs.output_primitive) {
-         case SHADER_PRIM_POINTS:
-            prim = ctx->gs_emitted_vertices;
-            break;
-         case SHADER_PRIM_LINE_STRIP:
-            prim = LLVMBuildSub(ctx->ac.builder, ctx->gs_emitted_vertices, ctx->ac.i32_1, "");
-            prim = ac_build_imax(&ctx->ac, prim, ctx->ac.i32_0);
-            break;
-         case SHADER_PRIM_TRIANGLE_STRIP:
-            prim = LLVMBuildSub(ctx->ac.builder, ctx->gs_emitted_vertices, LLVMConstInt(ctx->ac.i32, 2, 0), "");
-            prim = ac_build_imax(&ctx->ac, prim, ctx->ac.i32_0);
-            break;
-         }
-
-         LLVMValueRef args[] = {
-            prim,
-            ngg_get_emulated_counters_buf(ctx),
-            LLVMConstInt(ctx->ac.i32,
-                         si_query_pipestat_end_dw_offset(ctx->screen, PIPE_STAT_QUERY_GS_PRIMITIVES) * 4,
-                         false),
-            ctx->ac.i32_0,                            /* soffset */
-            ctx->ac.i32_0,                            /* cachepolicy */
-         };
-         ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32", ctx->ac.i32, args, 5, 0);
-
-         args[0] = ctx->ac.i32_1;
-         args[2] = LLVMConstInt(ctx->ac.i32,
-                                si_query_pipestat_end_dw_offset(ctx->screen, PIPE_STAT_QUERY_GS_INVOCATIONS) * 4,
-                                false);
-         ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32", ctx->ac.i32, args, 5, 0);
-      }
-      ac_build_endif(&ctx->ac, 5229);
-   }
-
    ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE, si_get_gs_wave_id(ctx));
 
    if (ctx->screen->info.gfx_level >= GFX9)
@@ -170,40 +122,9 @@ static void si_llvm_emit_vertex(struct ac_shader_abi *abi, unsigned stream,
 
    assert(!ctx->shader->key.ge.as_ngg);
 
-   struct si_shader_info *info = &ctx->shader->selector->info;
-   struct si_shader *shader = ctx->shader;
-   LLVMValueRef soffset = ac_get_arg(&ctx->ac, ctx->args->ac.gs2vs_offset);
-
-   unsigned offset = 0;
-   for (unsigned i = 0; i < info->num_outputs; i++) {
-      for (unsigned chan = 0; chan < 4; chan++) {
-         if (!(info->output_usagemask[i] & (1 << chan)) ||
-             ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
-            continue;
-
-         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++;
-
-         voffset = LLVMBuildAdd(ctx->ac.builder, voffset, vertexidx, "");
-         voffset = LLVMBuildMul(ctx->ac.builder, voffset, LLVMConstInt(ctx->ac.i32, 4, 0), "");
-
-         out_val = ac_to_integer(&ctx->ac, out_val);
-
-         ac_build_buffer_store_dword(&ctx->ac, ctx->gsvs_ring[stream], out_val, NULL,
-                                     voffset, soffset, ac_glc | ac_slc | ac_swizzled);
-      }
-   }
-
-   /* Signal vertex emission if vertex data was written. */
-   if (offset) {
-      ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
-                       si_get_gs_wave_id(ctx));
-
-      ctx->gs_emitted_vertices = LLVMBuildAdd(ctx->ac.builder, ctx->gs_emitted_vertices,
-                                              ctx->ac.i32_1, "vert");
-   }
+   /* Signal vertex emission */
+   ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
+                    si_get_gs_wave_id(ctx));
 }
 
 /* Cut one primitive from the geometry shader */