radeonsi/gfx10: implement streamout-related queries

[mesa.git] / src / gallium / drivers / radeonsi / gfx10_shader_ngg.c

/*
 * Copyright 2017 Advanced Micro Devices, Inc.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * on the rights to use, copy, modify, merge, publish, distribute, sub
 * license, and/or sell copies of the Software, and to permit persons to whom
 * the Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 */

#include "si_pipe.h"
#include "si_shader_internal.h"

#include "sid.h"

#include "util/u_memory.h"
#include "util/u_prim.h"

static LLVMValueRef get_wave_id_in_tg(struct si_shader_context *ctx)
{
        return si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
}

static LLVMValueRef get_tgsize(struct si_shader_context *ctx)
{
        return si_unpack_param(ctx, ctx->param_merged_wave_info, 28, 4);
}

static LLVMValueRef get_thread_id_in_tg(struct si_shader_context *ctx)
{
        LLVMBuilderRef builder = ctx->ac.builder;
        LLVMValueRef tmp;
        tmp = LLVMBuildMul(builder, get_wave_id_in_tg(ctx),
                           LLVMConstInt(ctx->ac.i32, 64, false), "");
        return LLVMBuildAdd(builder, tmp, ac_get_thread_id(&ctx->ac), "");
}

static LLVMValueRef ngg_get_vtx_cnt(struct si_shader_context *ctx)
{
        return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
                            LLVMConstInt(ctx->ac.i32, 12, false),
                            LLVMConstInt(ctx->ac.i32, 9, false),
                            false);
}

static LLVMValueRef ngg_get_prim_cnt(struct si_shader_context *ctx)
{
        return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
                            LLVMConstInt(ctx->ac.i32, 22, false),
                            LLVMConstInt(ctx->ac.i32, 9, false),
                            false);
}

static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
{
        LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
                                            ctx->param_rw_buffers);

        return ac_build_load_to_sgpr(&ctx->ac, buf_ptr,
                                     LLVMConstInt(ctx->i32, GFX10_GS_QUERY_BUF, false));
}

/* Send GS Alloc Req message from the first wave of the group to SPI.
 * Message payload is:
 * - bits 0..10: vertices in group
 * - bits 12..22: primitives in group
 */
static void build_sendmsg_gs_alloc_req(struct si_shader_context *ctx,
                                       LLVMValueRef vtx_cnt,
                                       LLVMValueRef prim_cnt)
{
        LLVMBuilderRef builder = ctx->ac.builder;
        LLVMValueRef tmp;

        tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
        ac_build_ifcc(&ctx->ac, tmp, 5020);

        tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->ac.i32, 12, false),"");
        tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
        ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_ALLOC_REQ, tmp);

        ac_build_endif(&ctx->ac, 5020);
}

struct ngg_prim {
        unsigned num_vertices;
        LLVMValueRef isnull;
        LLVMValueRef index[3];
        LLVMValueRef edgeflag[3];
};

static void build_export_prim(struct si_shader_context *ctx,
                              const struct ngg_prim *prim)
{
        LLVMBuilderRef builder = ctx->ac.builder;
        struct ac_export_args args;
        LLVMValueRef tmp;

        tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
        args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");

        for (unsigned i = 0; i < prim->num_vertices; ++i) {
                tmp = LLVMBuildShl(builder, prim->index[i],
                                   LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
                args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
                tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
                tmp = LLVMBuildShl(builder, tmp,
                                   LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
                args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
        }

        args.out[0] = LLVMBuildBitCast(builder, args.out[0], ctx->ac.f32, "");
        args.out[1] = LLVMGetUndef(ctx->ac.f32);
        args.out[2] = LLVMGetUndef(ctx->ac.f32);
        args.out[3] = LLVMGetUndef(ctx->ac.f32);

        args.target = V_008DFC_SQ_EXP_PRIM;
        args.enabled_channels = 1;
        args.done = true;
        args.valid_mask = false;
        args.compr = false;

        ac_build_export(&ctx->ac, &args);
}

/**
 * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
 */
void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
                             unsigned max_outputs,
                             LLVMValueRef *addrs)
{
        struct si_shader_context *ctx = si_shader_context_from_abi(abi);
        struct tgsi_shader_info *info = &ctx->shader->selector->info;
        struct si_shader_output_values *outputs = NULL;
        LLVMBuilderRef builder = ctx->ac.builder;
        struct lp_build_if_state if_state;
        LLVMValueRef tmp;

        assert(!ctx->shader->is_gs_copy_shader);
        assert(info->num_outputs <= max_outputs);

        outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));

        for (unsigned i = 0; i < info->num_outputs; i++) {
                outputs[i].semantic_name = info->output_semantic_name[i];
                outputs[i].semantic_index = info->output_semantic_index[i];

                /* This is used only by streamout. */
                for (unsigned j = 0; j < 4; j++) {
                        outputs[i].values[j] =
                                LLVMBuildLoad(builder,
                                              addrs[4 * i + j],
                                              "");
                        outputs[i].vertex_stream[j] =
                                (info->output_streams[i] >> (2 * j)) & 3;
                }
        }

        lp_build_endif(&ctx->merged_wrap_if_state);

        LLVMValueRef prims_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 8, 8);
        LLVMValueRef vtx_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 0, 8);
        LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
                                                  ac_get_thread_id(&ctx->ac), prims_in_wave, "");
        LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
                                                  ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
        LLVMValueRef vtxindex[] = {
                si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 0, 16),
                si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 16, 16),
                si_unpack_param(ctx, ctx->param_gs_vtx23_offset, 0, 16),
        };

        /* Determine the number of vertices per primitive. */
        unsigned num_vertices;
        LLVMValueRef num_vertices_val;

        if (ctx->type == PIPE_SHADER_VERTEX) {
                if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
                        /* Blits always use axis-aligned rectangles with 3 vertices. */
                        num_vertices = 3;
                        num_vertices_val = LLVMConstInt(ctx->i32, 3, 0);
                } else {
                        /* Extract OUTPRIM field. */
                        tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 2, 2);
                        num_vertices_val = LLVMBuildAdd(builder, tmp, ctx->i32_1, "");
                        num_vertices = 3; /* TODO: optimize for points & lines */
                }
        } else {
                assert(ctx->type == PIPE_SHADER_TESS_EVAL);

                if (info->properties[TGSI_PROPERTY_TES_POINT_MODE])
                        num_vertices = 1;
                else if (info->properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
                        num_vertices = 2;
                else
                        num_vertices = 3;

                num_vertices_val = LLVMConstInt(ctx->i32, num_vertices, false);
        }

        /* TODO: streamout */

        /* TODO: primitive culling */

        build_sendmsg_gs_alloc_req(ctx, ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));

        /* Update query buffer */
        tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
        ac_build_ifcc(&ctx->ac, tmp, 5030);
        tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac), ctx->ac.i32_0, "");
        ac_build_ifcc(&ctx->ac, tmp, 5031);
        {
                LLVMValueRef args[] = {
                        ngg_get_prim_cnt(ctx),
                        ngg_get_query_buf(ctx),
                        LLVMConstInt(ctx->i32, 16, false), /* offset of stream[0].generated_primitives */
                        ctx->i32_0, /* soffset */
                        ctx->i32_0, /* cachepolicy */
                };

                /* TODO: should this be 64-bit atomics? */
                ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
                                   ctx->i32, args, 5, 0);
        }
        ac_build_endif(&ctx->ac, 5031);
        ac_build_endif(&ctx->ac, 5030);

        /* Export primitive data to the index buffer. Format is:
         *  - bits 0..8: index 0
         *  - bit 9: edge flag 0
         *  - bits 10..18: index 1
         *  - bit 19: edge flag 1
         *  - bits 20..28: index 2
         *  - bit 29: edge flag 2
         *  - bit 31: null primitive (skip)
         *
         * For the first version, we will always build up all three indices
         * independent of the primitive type. The additional garbage data
         * shouldn't hurt.
         *
         * TODO: culling depends on the primitive type, so can have some
         * interaction here.
         */
        lp_build_if(&if_state, &ctx->gallivm, is_gs_thread);
        {
                struct ngg_prim prim = {};

                prim.num_vertices = num_vertices;
                prim.isnull = ctx->ac.i1false;
                memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);

                for (unsigned i = 0; i < num_vertices; ++i) {
                        tmp = LLVMBuildLShr(builder, ctx->abi.gs_invocation_id,
                                            LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
                        prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
                }

                build_export_prim(ctx, &prim);
        }
        lp_build_endif(&if_state);

        /* Export per-vertex data (positions and parameters). */
        lp_build_if(&if_state, &ctx->gallivm, is_es_thread);
        {
                unsigned i;

                /* Unconditionally (re-)load the values for proper SSA form. */
                for (i = 0; i < info->num_outputs; i++) {
                        for (unsigned j = 0; j < 4; j++) {
                                outputs[i].values[j] =
                                        LLVMBuildLoad(builder,
                                                addrs[4 * i + j],
                                                "");
                        }
                }

                /* TODO: Vertex shaders have to get PrimitiveID from GS VGPRs. */
                if (ctx->type == PIPE_SHADER_TESS_EVAL &&
                    ctx->shader->key.mono.u.vs_export_prim_id) {
                        outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
                        outputs[i].semantic_index = 0;
                        outputs[i].values[0] = ac_to_float(&ctx->ac, si_get_primitive_id(ctx, 0));
                        for (unsigned j = 1; j < 4; j++)
                                outputs[i].values[j] = LLVMGetUndef(ctx->f32);

                        memset(outputs[i].vertex_stream, 0,
                               sizeof(outputs[i].vertex_stream));
                        i++;
                }

                si_llvm_export_vs(ctx, outputs, i);
        }
        lp_build_endif(&if_state);

        FREE(outputs);
}

static LLVMValueRef
ngg_gs_get_vertex_storage(struct si_shader_context *ctx)
{
        const struct si_shader_selector *sel = ctx->shader->selector;
        const struct tgsi_shader_info *info = &sel->info;

        LLVMTypeRef elements[2] = {
                LLVMArrayType(ctx->ac.i32, 4 * info->num_outputs),
                LLVMArrayType(ctx->ac.i8, 4),
        };
        LLVMTypeRef type = LLVMStructTypeInContext(ctx->ac.context, elements, 2, false);
        type = LLVMPointerType(LLVMArrayType(type, 0), AC_ADDR_SPACE_LDS);
        return LLVMBuildBitCast(ctx->ac.builder, ctx->gs_ngg_emit, type, "");
}

/**
 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
 * is in emit order; that is:
 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
 * - during vertex emit, i.e. while the API GS shader invocation is running,
 *   N = threadidx * gs_max_out_vertices + emitidx
 *
 * Goals of the LDS memory layout:
 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
 *    in uniform control flow
 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
 *    culling
 * 3. Agnostic to the number of waves (since we don't know it before compiling)
 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
 * 5. Avoid wasting memory.
 *
 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
 * layout, elimination of bank conflicts requires that each vertex occupy an
 * odd number of dwords. We use the additional dword to store the output stream
 * index as well as a flag to indicate whether this vertex ends a primitive
 * for rasterization.
 *
 * Swizzling is required to satisfy points 1 and 2 simultaneously.
 *
 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
 * Indices are swizzled in groups of 32, which ensures point 1 without
 * disturbing point 2.
 *
 * \return an LDS pointer to type {[N x i32], [4 x i8]}
 */
static LLVMValueRef
ngg_gs_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef vertexidx)
{
        struct si_shader_selector *sel = ctx->shader->selector;
        LLVMBuilderRef builder = ctx->ac.builder;
        LLVMValueRef storage = ngg_gs_get_vertex_storage(ctx);

        /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
        unsigned write_stride_2exp = ffs(sel->gs_max_out_vertices) - 1;
        if (write_stride_2exp) {
                LLVMValueRef row =
                        LLVMBuildLShr(builder, vertexidx,
                                      LLVMConstInt(ctx->ac.i32, 5, false), "");
                LLVMValueRef swizzle =
                        LLVMBuildAnd(builder, row,
                                     LLVMConstInt(ctx->ac.i32, (1u << write_stride_2exp) - 1,
                                                  false), "");
                vertexidx = LLVMBuildXor(builder, vertexidx, swizzle, "");
        }

        return ac_build_gep0(&ctx->ac, storage, vertexidx);
}

static LLVMValueRef
ngg_gs_emit_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef gsthread,
                       LLVMValueRef emitidx)
{
        struct si_shader_selector *sel = ctx->shader->selector;
        LLVMBuilderRef builder = ctx->ac.builder;
        LLVMValueRef tmp;

        tmp = LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false);
        tmp = LLVMBuildMul(builder, tmp, gsthread, "");
        const LLVMValueRef vertexidx = LLVMBuildAdd(builder, tmp, emitidx, "");
        return ngg_gs_vertex_ptr(ctx, vertexidx);
}

void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx,
                              unsigned stream,
                              LLVMValueRef *addrs)
{
        const struct si_shader_selector *sel = ctx->shader->selector;
        const struct tgsi_shader_info *info = &sel->info;
        LLVMBuilderRef builder = ctx->ac.builder;
        struct lp_build_if_state if_state;
        LLVMValueRef tmp;
        const LLVMValueRef vertexidx =
                LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");

        /* If this thread has already emitted the declared maximum number of
         * vertices, skip the write: excessive vertex emissions are not
         * supposed to have any effect.
         */
        const LLVMValueRef can_emit =
                LLVMBuildICmp(builder, LLVMIntULT, vertexidx,
                              LLVMConstInt(ctx->i32, sel->gs_max_out_vertices, false), "");

        tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
        tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
        LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);

        lp_build_if(&if_state, &ctx->gallivm, can_emit);

        const LLVMValueRef vertexptr =
                ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
        unsigned out_idx = 0;
        for (unsigned i = 0; i < info->num_outputs; i++) {
                for (unsigned chan = 0; chan < 4; chan++, out_idx++) {
                        if (!(info->output_usagemask[i] & (1 << chan)) ||
                            ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
                                continue;

                        LLVMValueRef out_val = LLVMBuildLoad(builder, addrs[4 * i + chan], "");
                        LLVMValueRef gep_idx[3] = {
                                ctx->ac.i32_0, /* implied C-style array */
                                ctx->ac.i32_0, /* first entry of struct */
                                LLVMConstInt(ctx->ac.i32, out_idx, false),
                        };
                        LLVMValueRef ptr = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");

                        out_val = ac_to_integer(&ctx->ac, out_val);
                        LLVMBuildStore(builder, out_val, ptr);
                }
        }
        assert(out_idx * 4 == sel->gsvs_vertex_size);

        /* Determine and store whether this vertex completed a primitive. */
        const LLVMValueRef curverts = LLVMBuildLoad(builder, ctx->gs_curprim_verts[stream], "");

        tmp = LLVMConstInt(ctx->ac.i32, u_vertices_per_prim(sel->gs_output_prim) - 1, false);
        const LLVMValueRef iscompleteprim =
                LLVMBuildICmp(builder, LLVMIntUGE, curverts, tmp, "");

        tmp = LLVMBuildAdd(builder, curverts, ctx->ac.i32_1, "");
        LLVMBuildStore(builder, tmp, ctx->gs_curprim_verts[stream]);

        LLVMValueRef gep_idx[3] = {
                ctx->ac.i32_0, /* implied C-style array */
                ctx->ac.i32_1, /* second struct entry */
                LLVMConstInt(ctx->ac.i32, stream, false),
        };
        const LLVMValueRef primflagptr =
                LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");

        tmp = LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i8, "");
        LLVMBuildStore(builder, tmp, primflagptr);

        tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
        tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
        LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);

        lp_build_endif(&if_state);
}

void gfx10_ngg_gs_emit_prologue(struct si_shader_context *ctx)
{
        /* Zero out the part of LDS scratch that is used to accumulate the
         * per-stream generated primitive count.
         */
        LLVMBuilderRef builder = ctx->ac.builder;
        LLVMValueRef scratchptr = ctx->gs_ngg_scratch;
        LLVMValueRef tid = get_thread_id_in_tg(ctx);
        LLVMValueRef tmp;

        tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->i32, 4, false), "");
        ac_build_ifcc(&ctx->ac, tmp, 5090);
        {
                LLVMValueRef ptr = ac_build_gep0(&ctx->ac, scratchptr, tid);
                LLVMBuildStore(builder, ctx->i32_0, ptr);
        }
        ac_build_endif(&ctx->ac, 5090);

        ac_build_s_barrier(&ctx->ac);
}

void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
{
        const struct si_shader_selector *sel = ctx->shader->selector;
        const struct tgsi_shader_info *info = &sel->info;
        const unsigned verts_per_prim = u_vertices_per_prim(sel->gs_output_prim);
        LLVMBuilderRef builder = ctx->ac.builder;
        LLVMValueRef i8_0 = LLVMConstInt(ctx->ac.i8, 0, false);
        LLVMValueRef tmp, tmp2;

        /* Zero out remaining (non-emitted) primitive flags.
         *
         * Note: Alternatively, we could pass the relevant gs_next_vertex to
         *       the emit threads via LDS. This is likely worse in the expected
         *       typical case where each GS thread emits the full set of
         *       vertices.
         */
        for (unsigned stream = 0; stream < 4; ++stream) {
                if (!info->num_stream_output_components[stream])
                        continue;

                const LLVMValueRef gsthread = get_thread_id_in_tg(ctx);

                ac_build_bgnloop(&ctx->ac, 5100);

                const LLVMValueRef vertexidx =
                        LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
                tmp = LLVMBuildICmp(builder, LLVMIntUGE, vertexidx,
                        LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false), "");
                ac_build_ifcc(&ctx->ac, tmp, 5101);
                ac_build_break(&ctx->ac);
                ac_build_endif(&ctx->ac, 5101);

                tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
                LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);

                tmp = ngg_gs_emit_vertex_ptr(ctx, gsthread, vertexidx);
                LLVMValueRef gep_idx[3] = {
                        ctx->ac.i32_0, /* implied C-style array */
                        ctx->ac.i32_1, /* second entry of struct */
                        LLVMConstInt(ctx->ac.i32, stream, false),
                };
                tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
                LLVMBuildStore(builder, i8_0, tmp);

                ac_build_endloop(&ctx->ac, 5100);
        }

        /* Accumulate generated primitives counts across the entire threadgroup. */
        for (unsigned stream = 0; stream < 4; ++stream) {
                if (!info->num_stream_output_components[stream])
                        continue;

                LLVMValueRef numprims =
                        LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
                numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, 64);

                tmp = LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(&ctx->ac), ctx->i32_0, "");
                ac_build_ifcc(&ctx->ac, tmp, 5105);
                {
                        LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpAdd,
                                           ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
                                                         LLVMConstInt(ctx->i32, stream, false)),
                                           numprims, LLVMAtomicOrderingMonotonic, false);
                }
                ac_build_endif(&ctx->ac, 5105);
        }

        lp_build_endif(&ctx->merged_wrap_if_state);

        ac_build_s_barrier(&ctx->ac);

        const LLVMValueRef tid = get_thread_id_in_tg(ctx);
        LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);

        /* TODO: streamout */

        tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->i32, 4, false), "");
        ac_build_ifcc(&ctx->ac, tmp, 5110);
        {
                LLVMValueRef offset;
                tmp = tid;
                if (sel->so.num_outputs)
                        tmp = LLVMBuildAnd(builder, tmp, LLVMConstInt(ctx->i32, 3, false), "");
                offset = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 32, false), "");
                if (sel->so.num_outputs) {
                        tmp = LLVMBuildLShr(builder, tid, LLVMConstInt(ctx->i32, 2, false), "");
                        tmp = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 8, false), "");
                        offset = LLVMBuildAdd(builder, offset, tmp, "");
                }

                tmp = LLVMBuildLoad(builder, ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid), "");
                LLVMValueRef args[] = {
                        tmp,
                        ngg_get_query_buf(ctx),
                        offset,
                        LLVMConstInt(ctx->i32, 16, false), /* soffset */
                        ctx->i32_0, /* cachepolicy */
                };
                ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
                                   ctx->i32, args, 5, 0);
        }
        ac_build_endif(&ctx->ac, 5110);

        /* TODO: culling */

        /* Determine vertex liveness. */
        LLVMValueRef vertliveptr = lp_build_alloca(&ctx->gallivm, ctx->ac.i1, "vertexlive");

        tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
        ac_build_ifcc(&ctx->ac, tmp, 5120);
        {
                for (unsigned i = 0; i < verts_per_prim; ++i) {
                        const LLVMValueRef primidx =
                                LLVMBuildAdd(builder, tid,
                                             LLVMConstInt(ctx->ac.i32, i, false), "");

                        if (i > 0) {
                                tmp = LLVMBuildICmp(builder, LLVMIntULT, primidx, num_emit_threads, "");
                                ac_build_ifcc(&ctx->ac, tmp, 5121 + i);
                        }

                        /* Load primitive liveness */
                        tmp = ngg_gs_vertex_ptr(ctx, primidx);
                        LLVMValueRef gep_idx[3] = {
                                ctx->ac.i32_0, /* implicit C-style array */
                                ctx->ac.i32_1, /* second value of struct */
                                ctx->ac.i32_0, /* stream 0 */
                        };
                        tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
                        tmp = LLVMBuildLoad(builder, tmp, "");
                        const LLVMValueRef primlive =
                                LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");

                        tmp = LLVMBuildLoad(builder, vertliveptr, "");
                        tmp = LLVMBuildOr(builder, tmp, primlive, ""),
                        LLVMBuildStore(builder, tmp, vertliveptr);

                        if (i > 0)
                                ac_build_endif(&ctx->ac, 5121 + i);
                }
        }
        ac_build_endif(&ctx->ac, 5120);

        /* Inclusive scan addition across the current wave. */
        LLVMValueRef vertlive = LLVMBuildLoad(builder, vertliveptr, "");
        struct ac_wg_scan vertlive_scan = {};
        vertlive_scan.op = nir_op_iadd;
        vertlive_scan.enable_reduce = true;
        vertlive_scan.enable_exclusive = true;
        vertlive_scan.src = vertlive;
        vertlive_scan.scratch = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ctx->i32_0);
        vertlive_scan.waveidx = get_wave_id_in_tg(ctx);
        vertlive_scan.numwaves = get_tgsize(ctx);
        vertlive_scan.maxwaves = 8;

        ac_build_wg_scan(&ctx->ac, &vertlive_scan);

        /* Skip all exports (including index exports) when possible. At least on
         * early gfx10 revisions this is also to avoid hangs.
         */
        LLVMValueRef have_exports =
                LLVMBuildICmp(builder, LLVMIntNE, vertlive_scan.result_reduce, ctx->ac.i32_0, "");
        num_emit_threads =
                LLVMBuildSelect(builder, have_exports, num_emit_threads, ctx->ac.i32_0, "");

        /* Allocate export space. Send this message as early as possible, to
         * hide the latency of the SQ <-> SPI roundtrip.
         *
         * Note: We could consider compacting primitives for export as well.
         *       PA processes 1 non-null prim / clock, but it fetches 4 DW of
         *       prim data per clock and skips null primitives at no additional
         *       cost. So compacting primitives can only be beneficial when
         *       there are 4 or more contiguous null primitives in the export
         *       (in the common case of single-dword prim exports).
         */
        build_sendmsg_gs_alloc_req(ctx, vertlive_scan.result_reduce, num_emit_threads);

        /* Setup the reverse vertex compaction permutation. We re-use stream 1
         * of the primitive liveness flags, relying on the fact that each
         * threadgroup can have at most 256 threads. */
        ac_build_ifcc(&ctx->ac, vertlive, 5130);
        {
                tmp = ngg_gs_vertex_ptr(ctx, vertlive_scan.result_exclusive);
                LLVMValueRef gep_idx[3] = {
                        ctx->ac.i32_0, /* implicit C-style array */
                        ctx->ac.i32_1, /* second value of struct */
                        ctx->ac.i32_1, /* stream 1 */
                };
                tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
                tmp2 = LLVMBuildTrunc(builder, tid, ctx->ac.i8, "");
                LLVMBuildStore(builder, tmp2, tmp);
        }
        ac_build_endif(&ctx->ac, 5130);

        ac_build_s_barrier(&ctx->ac);

        /* Export primitive data */
        tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
        ac_build_ifcc(&ctx->ac, tmp, 5140);
        {
                struct ngg_prim prim = {};
                prim.num_vertices = verts_per_prim;

                tmp = ngg_gs_vertex_ptr(ctx, tid);
                LLVMValueRef gep_idx[3] = {
                        ctx->ac.i32_0, /* implicit C-style array */
                        ctx->ac.i32_1, /* second value of struct */
                        ctx->ac.i32_0, /* primflag */
                };
                tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
                tmp = LLVMBuildLoad(builder, tmp, "");
                prim.isnull = LLVMBuildICmp(builder, LLVMIntEQ, tmp,
                                            LLVMConstInt(ctx->ac.i8, 0, false), "");

                for (unsigned i = 0; i < verts_per_prim; ++i) {
                        prim.index[i] = LLVMBuildSub(builder, vertlive_scan.result_exclusive,
                                LLVMConstInt(ctx->ac.i32, verts_per_prim - i - 1, false), "");
                        prim.edgeflag[i] = ctx->ac.i1false;
                }

                build_export_prim(ctx, &prim);
        }
        ac_build_endif(&ctx->ac, 5140);

        /* Export position and parameter data */
        tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, vertlive_scan.result_reduce, "");
        ac_build_ifcc(&ctx->ac, tmp, 5145);
        {
                struct si_shader_output_values *outputs = NULL;
                outputs = MALLOC(info->num_outputs * sizeof(outputs[0]));

                tmp = ngg_gs_vertex_ptr(ctx, tid);
                LLVMValueRef gep_idx[3] = {
                        ctx->ac.i32_0, /* implicit C-style array */
                        ctx->ac.i32_1, /* second value of struct */
                        ctx->ac.i32_1, /* stream 1: source data index */
                };
                tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
                tmp = LLVMBuildLoad(builder, tmp, "");
                tmp = LLVMBuildZExt(builder, tmp, ctx->ac.i32, "");
                const LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tmp);

                unsigned out_idx = 0;
                gep_idx[1] = ctx->ac.i32_0;
                for (unsigned i = 0; i < info->num_outputs; i++) {
                        outputs[i].semantic_name = info->output_semantic_name[i];
                        outputs[i].semantic_index = info->output_semantic_index[i];

                        for (unsigned j = 0; j < 4; j++, out_idx++) {
                                gep_idx[2] = LLVMConstInt(ctx->ac.i32, out_idx, false);
                                tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
                                tmp = LLVMBuildLoad(builder, tmp, "");
                                outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
                                outputs[i].vertex_stream[j] =
                                        (info->output_streams[i] >> (2 * j)) & 3;
                        }
                }

                si_llvm_export_vs(ctx, outputs, info->num_outputs);

                FREE(outputs);
        }
        ac_build_endif(&ctx->ac, 5145);
}

static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts,
                                     unsigned min_verts_per_prim, bool use_adjacency)
{
        unsigned max_reuse = max_esverts - min_verts_per_prim;
        if (use_adjacency)
                max_reuse /= 2;
        *max_gsprims = MIN2(*max_gsprims, 1 + max_reuse);
}

/**
 * Determine subgroup information like maximum number of vertices and prims.
 *
 * This happens before the shader is uploaded, since LDS relocations during
 * upload depend on the subgroup size.
 */
void gfx10_ngg_calculate_subgroup_info(struct si_shader *shader)
{
        const struct si_shader_selector *gs_sel = shader->selector;
        const struct si_shader_selector *es_sel =
                shader->previous_stage_sel ? shader->previous_stage_sel : gs_sel;
        const enum pipe_shader_type gs_type = gs_sel->type;
        const unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
        /* TODO: Specialize for known primitive type without GS. */
        const unsigned input_prim = gs_type == PIPE_SHADER_GEOMETRY ?
                                    gs_sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM] :
                                    PIPE_PRIM_TRIANGLES;
        const bool use_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
                                   input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
        const unsigned max_verts_per_prim = u_vertices_per_prim(input_prim);
        const unsigned min_verts_per_prim =
                gs_type == PIPE_SHADER_GEOMETRY ? max_verts_per_prim : 1;

        /* All these are in dwords: */
        /* We can't allow using the whole LDS, because GS waves compete with
         * other shader stages for LDS space.
         *
         * Streamout can increase the ESGS buffer size later on, so be more
         * conservative with streamout and use 4K dwords. This may be suboptimal.
         *
         * Otherwise, use the limit of 7K dwords. The reason is that we need
         * to leave some headroom for the max_esverts increase at the end.
         *
         * TODO: We should really take the shader's internal LDS use into
         *       account. The linker will fail if the size is greater than
         *       8K dwords.
         */
        const unsigned max_lds_size = (gs_sel->so.num_outputs ? 4 : 7) * 1024 - 128;
        const unsigned target_lds_size = max_lds_size;
        unsigned esvert_lds_size = 0;
        unsigned gsprim_lds_size = 0;

        /* All these are per subgroup: */
        bool max_vert_out_per_gs_instance = false;
        unsigned max_esverts_base = 256;
        unsigned max_gsprims_base = 128; /* default prim group size clamp */

        /* Hardware has the following non-natural restrictions on the value
         * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
         * the draw:
         *  - at most 252 for any line input primitive type
         *  - at most 251 for any quad input primitive type
         *  - at most 251 for triangle strips with adjacency (this happens to
         *    be the natural limit for triangle *lists* with adjacency)
         */
        max_esverts_base = MIN2(max_esverts_base, 251 + max_verts_per_prim - 1);

        if (gs_type == PIPE_SHADER_GEOMETRY) {
                unsigned max_out_verts_per_gsprim =
                        gs_sel->gs_max_out_vertices * gs_num_invocations;

                if (max_out_verts_per_gsprim <= 256) {
                        if (max_out_verts_per_gsprim) {
                                max_gsprims_base = MIN2(max_gsprims_base,
                                                        256 / max_out_verts_per_gsprim);
                        }
                } else {
                        /* Use special multi-cycling mode in which each GS
                         * instance gets its own subgroup. Does not work with
                         * tessellation. */
                        max_vert_out_per_gs_instance = true;
                        max_gsprims_base = 1;
                        max_out_verts_per_gsprim = gs_sel->gs_max_out_vertices;
                }

                esvert_lds_size = es_sel->esgs_itemsize / 4;
                gsprim_lds_size = (gs_sel->gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
        } else {
                /* TODO: This needs to be adjusted once LDS use for compaction
                 * after culling is implemented. */
        }

        unsigned max_gsprims = max_gsprims_base;
        unsigned max_esverts = max_esverts_base;

        if (esvert_lds_size)
                max_esverts = MIN2(max_esverts, target_lds_size / esvert_lds_size);
        if (gsprim_lds_size)
                max_gsprims = MIN2(max_gsprims, target_lds_size / gsprim_lds_size);

        max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
        clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, use_adjacency);
        assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);

        if (esvert_lds_size || gsprim_lds_size) {
                /* Now that we have a rough proportionality between esverts
                 * and gsprims based on the primitive type, scale both of them
                 * down simultaneously based on required LDS space.
                 *
                 * We could be smarter about this if we knew how much vertex
                 * reuse to expect.
                 */
                unsigned lds_total = max_esverts * esvert_lds_size +
                                     max_gsprims * gsprim_lds_size;
                if (lds_total > target_lds_size) {
                        max_esverts = max_esverts * target_lds_size / lds_total;
                        max_gsprims = max_gsprims * target_lds_size / lds_total;

                        max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
                        clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
                                                 min_verts_per_prim, use_adjacency);
                        assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
                }
        }

        /* Round up towards full wave sizes for better ALU utilization. */
        if (!max_vert_out_per_gs_instance) {
                const unsigned wavesize = 64;
                unsigned orig_max_esverts;
                unsigned orig_max_gsprims;
                do {
                        orig_max_esverts = max_esverts;
                        orig_max_gsprims = max_gsprims;

                        max_esverts = align(max_esverts, wavesize);
                        max_esverts = MIN2(max_esverts, max_esverts_base);
                        if (esvert_lds_size)
                                max_esverts = MIN2(max_esverts,
                                                   (max_lds_size - max_gsprims * gsprim_lds_size) /
                                                   esvert_lds_size);
                        max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);

                        max_gsprims = align(max_gsprims, wavesize);
                        max_gsprims = MIN2(max_gsprims, max_gsprims_base);
                        if (gsprim_lds_size)
                                max_gsprims = MIN2(max_gsprims,
                                                   (max_lds_size - max_esverts * esvert_lds_size) /
                                                   gsprim_lds_size);
                        clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
                                                 min_verts_per_prim, use_adjacency);
                        assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
                } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims);
        }

        /* Hardware restriction: minimum value of max_esverts */
        max_esverts = MAX2(max_esverts, 23 + max_verts_per_prim);

        unsigned max_out_vertices =
                max_vert_out_per_gs_instance ? gs_sel->gs_max_out_vertices :
                gs_type == PIPE_SHADER_GEOMETRY ?
                max_gsprims * gs_num_invocations * gs_sel->gs_max_out_vertices :
                max_esverts;
        assert(max_out_vertices <= 256);

        unsigned prim_amp_factor = 1;
        if (gs_type == PIPE_SHADER_GEOMETRY) {
                /* Number of output primitives per GS input primitive after
                 * GS instancing. */
                prim_amp_factor = gs_sel->gs_max_out_vertices;
        }

        /* The GE only checks against the maximum number of ES verts after
         * allocating a full GS primitive. So we need to ensure that whenever
         * this check passes, there is enough space for a full primitive without
         * vertex reuse.
         */
        shader->ngg.hw_max_esverts = max_esverts - max_verts_per_prim + 1;
        shader->ngg.max_gsprims = max_gsprims;
        shader->ngg.max_out_verts = max_out_vertices;
        shader->ngg.prim_amp_factor = prim_amp_factor;
        shader->ngg.max_vert_out_per_gs_instance = max_vert_out_per_gs_instance;

        shader->gs_info.esgs_ring_size = 4 * max_esverts * esvert_lds_size;
        shader->ngg.ngg_emit_size = max_gsprims * gsprim_lds_size;

        assert(shader->ngg.hw_max_esverts >= 24); /* HW limitation */
}