From: Nicolai Hähnle <nicolai.haeh...@amd.com>

---
 src/amd/common/ac_nir_to_llvm.c | 28 ++++++++++++++++++++--------
 src/amd/common/ac_shader_abi.h  | 10 ++++++++++
 2 files changed, 30 insertions(+), 8 deletions(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 4fad709..d2238c3 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -2189,24 +2189,22 @@ static void visit_store_ssbo(struct nir_to_llvm_context 
*ctx,
 {
        const char *store_name;
        LLVMValueRef src_data = get_src(ctx->nir, instr->src[0]);
        LLVMTypeRef data_type = ctx->f32;
        int elem_size_mult = get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 32;
        int components_32bit = elem_size_mult * instr->num_components;
        unsigned writemask = nir_intrinsic_write_mask(instr);
        LLVMValueRef base_data, base_offset;
        LLVMValueRef params[6];
 
-       if (ctx->stage == MESA_SHADER_FRAGMENT)
-               ctx->shader_info->fs.writes_memory = true;
-
-       params[1] = get_src(ctx->nir, instr->src[1]);
+       params[1] = ctx->abi.load_ssbo(&ctx->abi,
+                                      get_src(ctx->nir, instr->src[1]), true);
        params[2] = LLVMConstInt(ctx->i32, 0, false); /* vindex */
        params[4] = ctx->i1false;  /* glc */
        params[5] = ctx->i1false;  /* slc */
 
        if (components_32bit > 1)
                data_type = LLVMVectorType(ctx->f32, components_32bit);
 
        base_data = to_float(&ctx->ac, src_data);
        base_data = trim_vector(&ctx->ac, base_data, instr->num_components);
        base_data = LLVMBuildBitCast(ctx->builder, base_data,
@@ -2269,28 +2267,28 @@ static void visit_store_ssbo(struct nir_to_llvm_context 
*ctx,
                                   ctx->voidt, params, 6, 0);
        }
 }
 
 static LLVMValueRef visit_atomic_ssbo(struct nir_to_llvm_context *ctx,
                                       const nir_intrinsic_instr *instr)
 {
        const char *name;
        LLVMValueRef params[6];
        int arg_count = 0;
-       if (ctx->stage == MESA_SHADER_FRAGMENT)
-               ctx->shader_info->fs.writes_memory = true;
 
        if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
                params[arg_count++] = llvm_extract_elem(&ctx->ac, 
get_src(ctx->nir, instr->src[3]), 0);
        }
        params[arg_count++] = llvm_extract_elem(&ctx->ac, get_src(ctx->nir, 
instr->src[2]), 0);
-       params[arg_count++] = get_src(ctx->nir, instr->src[0]);
+       params[arg_count++] = ctx->abi.load_ssbo(&ctx->abi,
+                                                get_src(ctx->nir, 
instr->src[0]),
+                                                true);
        params[arg_count++] = LLVMConstInt(ctx->i32, 0, false); /* vindex */
        params[arg_count++] = get_src(ctx->nir, instr->src[1]);      /* voffset 
*/
        params[arg_count++] = ctx->i1false;  /* slc */
 
        switch (instr->intrinsic) {
        case nir_intrinsic_ssbo_atomic_add:
                name = "llvm.amdgcn.buffer.atomic.add";
                break;
        case nir_intrinsic_ssbo_atomic_imin:
                name = "llvm.amdgcn.buffer.atomic.smin";
@@ -2350,21 +2348,23 @@ static LLVMValueRef visit_load_buffer(struct 
nir_to_llvm_context *ctx,
                if (load_components >= 3)
                        load_name = "llvm.amdgcn.buffer.load.v4f32";
                else if (load_components == 2)
                        load_name = "llvm.amdgcn.buffer.load.v2f32";
                else if (load_components == 1)
                        load_name = "llvm.amdgcn.buffer.load.f32";
                else
                        unreachable("unhandled number of components");
 
                LLVMValueRef params[] = {
-                       get_src(ctx->nir, instr->src[0]),
+                       ctx->abi.load_ssbo(&ctx->abi,
+                                          get_src(ctx->nir, instr->src[0]),
+                                          false),
                        LLVMConstInt(ctx->i32, 0, false),
                        offset,
                        ctx->i1false,
                        ctx->i1false,
                };
 
                results[i] = ac_build_intrinsic(&ctx->ac, load_name, data_type, 
params, 5, 0);
 
        }
 
@@ -4029,20 +4029,31 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
                fprintf(stderr, "Unknown intrinsic: ");
                nir_print_instr(&instr->instr, stderr);
                fprintf(stderr, "\n");
                break;
        }
        if (result) {
                _mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
        }
 }
 
+static LLVMValueRef radv_load_ssbo(struct ac_shader_abi *abi,
+                                  LLVMValueRef buffer, bool write)
+{
+       struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
+
+       if (write && ctx->stage == MESA_SHADER_FRAGMENT)
+               ctx->shader_info->fs.writes_memory = true;
+
+       return buffer;
+}
+
 static LLVMValueRef radv_get_sampler_desc(struct ac_shader_abi *abi,
                                          unsigned descriptor_set,
                                          unsigned base_index,
                                          unsigned constant_index,
                                          LLVMValueRef index,
                                          enum ac_descriptor_type desc_type,
                                          bool image, bool write)
 {
        struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
        LLVMValueRef list = ctx->descriptor_sets[descriptor_set];
@@ -6075,20 +6086,21 @@ LLVMModuleRef 
ac_translate_nir_to_llvm(LLVMTargetMachineRef tm,
 
        nir_foreach_variable(variable, &nir->inputs)
                handle_shader_input_decl(&ctx, variable);
 
        if (nir->stage == MESA_SHADER_FRAGMENT)
                handle_fs_inputs_pre(&ctx, nir);
 
        ctx.abi.chip_class = options->chip_class;
        ctx.abi.inputs = &ctx.inputs[0];
        ctx.abi.emit_outputs = handle_shader_outputs_post;
+       ctx.abi.load_ssbo = radv_load_ssbo;
        ctx.abi.load_sampler_desc = radv_get_sampler_desc;
 
        nir_foreach_variable(variable, &nir->outputs)
                scan_shader_output_decl(&ctx, variable);
 
        ac_nir_translate(&ctx.ac, &ctx.abi, nir, &ctx);
 
        LLVMBuildRetVoid(ctx.builder);
 
        ac_llvm_finalize_module(&ctx);
diff --git a/src/amd/common/ac_shader_abi.h b/src/amd/common/ac_shader_abi.h
index d3c1ee1..f9a2c2b 100644
--- a/src/amd/common/ac_shader_abi.h
+++ b/src/amd/common/ac_shader_abi.h
@@ -50,20 +50,30 @@ struct ac_shader_abi {
         */
        LLVMValueRef *inputs;
 
        void (*emit_outputs)(struct ac_shader_abi *abi,
                             unsigned max_outputs,
                             LLVMValueRef *addrs);
 
        LLVMValueRef (*load_ubo)(struct ac_shader_abi *abi, LLVMValueRef index);
 
        /**
+        * Load the descriptor for the given buffer.
+        *
+        * \param buffer the buffer as presented in NIR: this is the descriptor
+        *               in Vulkan, and the buffer index in OpenGL/Gallium
+        * \param write whether buffer contents will be written
+        */
+       LLVMValueRef (*load_ssbo)(struct ac_shader_abi *abi,
+                                 LLVMValueRef buffer, bool write);
+
+       /**
         * Load a descriptor associated to a sampler.
         *
         * \param descriptor_set the descriptor set index (only for Vulkan)
         * \param base_index the base index of the sampler variable
         * \param constant_index constant part of an array index (or 0, if the
         *                       sampler variable is not an array)
         * \param index non-constant part of an array index (may be NULL)
         * \param desc_type the type of descriptor to load
         * \param image whether the descriptor is loaded for an image operation
         */
-- 
2.9.3

_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

Reply via email to