radv: emit stream outputs for vertex and tessellation stages
authorSamuel Pitoiset <samuel.pitoiset@gmail.com>
Fri, 5 Oct 2018 15:54:22 +0000 (17:54 +0200)
committerSamuel Pitoiset <samuel.pitoiset@gmail.com>
Mon, 29 Oct 2018 16:09:08 +0000 (17:09 +0100)
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
src/amd/vulkan/radv_nir_to_llvm.c

index 0ebd3bbd34a0a82052d35b0844bfbc04b8517c70..46c96dfac06312137c33752491865cce6edc83bc 100644 (file)
@@ -2492,6 +2492,140 @@ radv_load_output(struct radv_shader_context *ctx, unsigned index, unsigned chan)
        return LLVMBuildLoad(ctx->ac.builder, output, "");
 }
 
+static void
+radv_emit_stream_output(struct radv_shader_context *ctx,
+                        LLVMValueRef const *so_buffers,
+                        LLVMValueRef const *so_write_offsets,
+                        const struct radv_stream_output *output)
+{
+       unsigned num_comps = util_bitcount(output->component_mask);
+       unsigned loc = output->location;
+       unsigned buf = output->buffer;
+       unsigned offset = output->offset;
+       unsigned start;
+       LLVMValueRef out[4];
+
+       assert(num_comps && num_comps <= 4);
+       if (!num_comps || num_comps > 4)
+               return;
+
+       /* Get the first component. */
+       start = ffs(output->component_mask) - 1;
+
+       /* Adjust the destination offset. */
+       offset += start * 4;
+
+       /* Load the output as int. */
+       for (int i = 0; i < num_comps; i++) {
+               out[i] = ac_to_integer(&ctx->ac,
+                                      radv_load_output(ctx, loc, start + i));
+       }
+
+       /* Pack the output. */
+       LLVMValueRef vdata = NULL;
+
+       switch (num_comps) {
+       case 1: /* as i32 */
+               vdata = out[0];
+               break;
+       case 2: /* as v2i32 */
+       case 3: /* as v4i32 (aligned to 4) */
+               out[3] = LLVMGetUndef(ctx->ac.i32);
+               /* fall through */
+       case 4: /* as v4i32 */
+               vdata = ac_build_gather_values(&ctx->ac, out,
+                                              util_next_power_of_two(num_comps));
+               break;
+       }
+
+       ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf],
+                                   vdata, num_comps, so_write_offsets[buf],
+                                   ctx->ac.i32_0, offset,
+                                   1, 1, true, false);
+}
+
+static void
+radv_emit_streamout(struct radv_shader_context *ctx, unsigned stream)
+{
+       struct ac_build_if_state if_ctx;
+       int i;
+
+       /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
+       assert(ctx->streamout_config);
+       LLVMValueRef so_vtx_count =
+               ac_build_bfe(&ctx->ac, ctx->streamout_config,
+                            LLVMConstInt(ctx->ac.i32, 16, false),
+                            LLVMConstInt(ctx->ac.i32, 7, false), false);
+
+       LLVMValueRef tid = ac_get_thread_id(&ctx->ac);
+
+       /* can_emit = tid < so_vtx_count; */
+       LLVMValueRef can_emit = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
+                                             tid, so_vtx_count, "");
+
+       /* Emit the streamout code conditionally. This actually avoids
+        * out-of-bounds buffer access. The hw tells us via the SGPR
+        * (so_vtx_count) which threads are allowed to emit streamout data.
+        */
+       ac_nir_build_if(&if_ctx, ctx, can_emit);
+       {
+               /* The buffer offset is computed as follows:
+                *   ByteOffset = streamout_offset[buffer_id]*4 +
+                *                (streamout_write_index + thread_id)*stride[buffer_id] +
+                *                attrib_offset
+                */
+               LLVMValueRef so_write_index = ctx->streamout_write_idx;
+
+               /* Compute (streamout_write_index + thread_id). */
+               so_write_index =
+                       LLVMBuildAdd(ctx->ac.builder, so_write_index, tid, "");
+
+               /* Load the descriptor and compute the write offset for each
+                * enabled buffer.
+                */
+               LLVMValueRef so_write_offset[4] = {};
+               LLVMValueRef so_buffers[4] = {};
+               LLVMValueRef buf_ptr = ctx->streamout_buffers;
+
+               for (i = 0; i < 4; i++) {
+                       uint16_t stride = ctx->shader_info->info.so.strides[i];
+
+                       if (!stride)
+                               continue;
+
+                       LLVMValueRef offset =
+                               LLVMConstInt(ctx->ac.i32, i, false);
+
+                       so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac,
+                                                             buf_ptr, offset);
+
+                       LLVMValueRef so_offset = ctx->streamout_offset[i];
+
+                       so_offset = LLVMBuildMul(ctx->ac.builder, so_offset,
+                                                LLVMConstInt(ctx->ac.i32, 4, false), "");
+
+                       so_write_offset[i] =
+                               ac_build_imad(&ctx->ac, so_write_index,
+                                             LLVMConstInt(ctx->ac.i32,
+                                                          stride * 4, false),
+                                             so_offset);
+               }
+
+               /* Write streamout data. */
+               for (i = 0; i < ctx->shader_info->info.so.num_outputs; i++) {
+                       struct radv_stream_output *output =
+                               &ctx->shader_info->info.so.outputs[i];
+
+                       if (stream != output->stream)
+                               continue;
+
+                       radv_emit_stream_output(ctx, so_buffers,
+                                               so_write_offset, output);
+               }
+       }
+       ac_nir_build_endif(&if_ctx);
+}
+
 static void
 handle_vs_outputs_post(struct radv_shader_context *ctx,
                       bool export_prim_id, bool export_layer_id,
@@ -2589,6 +2723,9 @@ handle_vs_outputs_post(struct radv_shader_context *ctx,
                viewport_index_value = radv_load_output(ctx, VARYING_SLOT_VIEWPORT, 0);
        }
 
+       if (ctx->shader_info->info.so.num_outputs)
+               radv_emit_streamout(ctx, 0);
+
        if (outinfo->writes_pointsize ||
            outinfo->writes_layer ||
            outinfo->writes_viewport_index) {