gallivm/llvmpipe: add support for block size intrinsic

We have to pass the main block size into the coroutine
and into the shader.

Acked-by: Roland Scheidegger <sroland@vmware.com>

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_nir.c b/src/gallium/auxiliary/gallivm/lp_bld_nir.c
index 82b847d05c9db7263ee2488045520a0046f9ebff..9c378d174d81a38f3d332b8f45f05187e06caa41 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_nir.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_nir.c
@@ -1246,6 +1246,7 @@ static void visit_intrinsic(struct lp_build_nir_context *bld_base,
    case nir_intrinsic_load_invocation_id:
    case nir_intrinsic_load_front_face:
    case nir_intrinsic_load_draw_id:
+   case nir_intrinsic_load_local_group_size:
    case nir_intrinsic_load_work_dim:
       bld_base->sysval_intrin(bld_base, instr, result);
       break;

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_nir_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_nir_soa.c
index 26183dba7236b60c136a3f3ae5ecdc9b9e902f66..99f49a942fe71d3e66628d9ae19e9d1d38489c9d 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_nir_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_nir_soa.c
@@ -996,6 +996,10 @@ static void emit_sysval_intrin(struct lp_build_nir_context *bld_base,
       break;
    default:
       break;
+   case nir_intrinsic_load_local_group_size:
+     for (unsigned i = 0; i < 3; i++)
+       result[i] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildExtractElement(gallivm->builder, bld->system_values.block_size, lp_build_const_int32(gallivm, i), ""));
+     break;
    case nir_intrinsic_load_work_dim:
       result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.work_dim);
       break;

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
index 3be4401838316d812e4218f9116d9b73d4a577a3..cb8c91eadd2b925ba329c81ca84f1ba6d6fcf005 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h
@@ -177,6 +177,7 @@ struct lp_bld_tgsi_system_values {
    LLVMValueRef grid_size;
    LLVMValueRef front_facing;
    LLVMValueRef work_dim;
+   LLVMValueRef block_size;
 };
 
 

diff --git a/src/gallium/drivers/llvmpipe/lp_state_cs.c b/src/gallium/drivers/llvmpipe/lp_state_cs.c
index f40336adecb95cbb3f07564105e0019f674c6f41..83876dd98d4e8182d8b91bac915510a3b6901752 100644 (file)
--- a/src/gallium/drivers/llvmpipe/lp_state_cs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_cs.c
@@ -63,7 +63,7 @@ generate_compute(struct llvmpipe_context *lp,
    struct gallivm_state *gallivm = variant->gallivm;
    const struct lp_compute_shader_variant_key *key = &variant->key;
    char func_name[64], func_name_coro[64];
-   LLVMTypeRef arg_types[14];
+   LLVMTypeRef arg_types[17];
    LLVMTypeRef func_type, coro_func_type;
    LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
    LLVMValueRef context_ptr;
@@ -112,6 +112,9 @@ generate_compute(struct llvmpipe_context *lp,
    arg_types[11] = variant->jit_cs_thread_data_ptr_type;  /* per thread data */
    arg_types[12] = int32_type;                         /* coro only - num X loops */
    arg_types[13] = int32_type;                         /* coro only - partials */
+   arg_types[14] = int32_type;                         /* coro block_x_size */
+   arg_types[15] = int32_type;                         /* coro block_y_size */
+   arg_types[16] = int32_type;                         /* coro block_z_size */
    func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
                                 arg_types, ARRAY_SIZE(arg_types) - 5, 0);
 
@@ -197,7 +200,7 @@ generate_compute(struct llvmpipe_context *lp,
    lp_build_loop_begin(&loop_state[0], gallivm,
                        lp_build_const_int32(gallivm, 0)); /* x loop */
    {
-      LLVMValueRef args[14];
+      LLVMValueRef args[17];
       args[0] = context_ptr;
       args[1] = loop_state[0].counter;
       args[2] = loop_state[1].counter;
@@ -212,6 +215,9 @@ generate_compute(struct llvmpipe_context *lp,
       args[11] = thread_data_ptr;
       args[12] = num_x_loop;
       args[13] = partials;
+      args[14] = x_size_arg;
+      args[15] = y_size_arg;
+      args[16] = z_size_arg;
 
       /* idx = (z * (size_x * size_y) + y * size_x + x */
       LLVMValueRef coro_hdl_idx = LLVMBuildMul(gallivm->builder, loop_state[2].counter,
@@ -231,7 +237,7 @@ generate_compute(struct llvmpipe_context *lp,
                                        lp_build_const_int32(gallivm, 0), "");
       /* first time here - call the coroutine function entry point */
       lp_build_if(&ifstate, gallivm, cmp);
-      LLVMValueRef coro_ret = LLVMBuildCall(gallivm->builder, coro, args, 14, "");
+      LLVMValueRef coro_ret = LLVMBuildCall(gallivm->builder, coro, args, 17, "");
       LLVMBuildStore(gallivm->builder, coro_ret, coro_entry);
       lp_build_else(&ifstate);
       /* subsequent calls for this invocation - check if done. */
@@ -263,6 +269,7 @@ generate_compute(struct llvmpipe_context *lp,
    LLVMBuildRetVoid(builder);
 
    /* This is stage (b) - generate the compute shader code inside the coroutine. */
+   LLVMValueRef block_x_size_arg, block_y_size_arg, block_z_size_arg;
    context_ptr  = LLVMGetParam(coro, 0);
    x_size_arg = LLVMGetParam(coro, 1);
    y_size_arg = LLVMGetParam(coro, 2);
@@ -277,6 +284,9 @@ generate_compute(struct llvmpipe_context *lp,
    thread_data_ptr  = LLVMGetParam(coro, 11);
    num_x_loop = LLVMGetParam(coro, 12);
    partials = LLVMGetParam(coro, 13);
+   block_x_size_arg = LLVMGetParam(coro, 14);
+   block_y_size_arg = LLVMGetParam(coro, 15);
+   block_z_size_arg = LLVMGetParam(coro, 16);
    block = LLVMAppendBasicBlockInContext(gallivm->context, coro, "entry");
    LLVMPositionBuilderAtEnd(builder, block);
    {
@@ -328,6 +338,11 @@ generate_compute(struct llvmpipe_context *lp,
 
       system_values.work_dim = work_dim_arg;
 
+      LLVMValueRef bsize[3] = { block_x_size_arg, block_y_size_arg, block_z_size_arg };
+      system_values.block_size = LLVMGetUndef(LLVMVectorType(int32_type, 3));
+      for (i = 0; i < 3; i++)
+         system_values.block_size = LLVMBuildInsertElement(builder, system_values.block_size, bsize[i], lp_build_const_int32(gallivm, i), "");
+
       LLVMValueRef last_x_loop = LLVMBuildICmp(gallivm->builder, LLVMIntEQ, x_size_arg, LLVMBuildSub(gallivm->builder, num_x_loop, lp_build_const_int32(gallivm, 1), ""), "");
       LLVMValueRef use_partial_mask = LLVMBuildAnd(gallivm->builder, last_x_loop, has_partials, "");
       struct lp_build_if_state if_state;