[Mesa-dev] [PATCH] radeonsi: only use new versions of LLVM image and sample intrinsics

Tom Stellard tom at stellard.net
Thu Sep 3 10:06:39 PDT 2015


On Thu, Sep 03, 2015 at 06:28:07PM +0200, Marek Olšák wrote:
> Sadly, this patch requires LLVM 3.5, which I didn't realize. Would it
> be okay to remove LLVM 3.4.2 support?
> 

Yes, that would be fine.

-Tom

> Marek
> 
> On Wed, Sep 2, 2015 at 7:48 PM, Tom Stellard <tom at stellard.net> wrote:
> > On Tue, Sep 01, 2015 at 11:18:15PM +0200, Marek Olšák wrote:
> >> From: Marek Olšák <marek.olsak at amd.com>
> >>
> >
> > Reviewed-by: Tom Stellard <thomas.stellard at amd.com>
> >> Just a cleanup I had made a long time ago and forgot about.
> >> ---
> >>  src/gallium/drivers/radeonsi/si_shader.c | 468 +++++++++++++------------------
> >>  1 file changed, 191 insertions(+), 277 deletions(-)
> >>
> >> diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
> >> index ab5b3ee..7550a42 100644
> >> --- a/src/gallium/drivers/radeonsi/si_shader.c
> >> +++ b/src/gallium/drivers/radeonsi/si_shader.c
> >> @@ -2256,6 +2256,64 @@ static bool tgsi_is_shadow_sampler(unsigned target)
> >>              target == TGSI_TEXTURE_SHADOWRECT;
> >>  }
> >>
> >> +static bool tgsi_is_array_sampler(unsigned target)
> >> +{
> >> +     return target == TGSI_TEXTURE_1D_ARRAY ||
> >> +            target == TGSI_TEXTURE_SHADOW1D_ARRAY ||
> >> +            target == TGSI_TEXTURE_2D_ARRAY ||
> >> +            target == TGSI_TEXTURE_SHADOW2D_ARRAY ||
> >> +            target == TGSI_TEXTURE_CUBE_ARRAY ||
> >> +            target == TGSI_TEXTURE_SHADOWCUBE_ARRAY ||
> >> +            target == TGSI_TEXTURE_2D_ARRAY_MSAA;
> >> +}
> >> +
> >> +static void set_tex_fetch_args(struct gallivm_state *gallivm,
> >> +                            struct lp_build_emit_data *emit_data,
> >> +                            unsigned opcode, unsigned target,
> >> +                            LLVMValueRef res_ptr, LLVMValueRef samp_ptr,
> >> +                            LLVMValueRef *param, unsigned count,
> >> +                            unsigned dmask)
> >> +{
> >> +     unsigned num_args;
> >> +     unsigned is_rect = target == TGSI_TEXTURE_RECT;
> >> +     LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
> >> +
> >> +     /* Pad to power of two vector */
> >> +     while (count < util_next_power_of_two(count))
> >> +             param[count++] = LLVMGetUndef(i32);
> >> +
> >> +     /* Texture coordinates. */
> >> +     if (count > 1)
> >> +             emit_data->args[0] = lp_build_gather_values(gallivm, param, count);
> >> +     else
> >> +             emit_data->args[0] = param[0];
> >> +
> >> +     /* Resource. */
> >> +     emit_data->args[1] = res_ptr;
> >> +     num_args = 2;
> >> +
> >> +     if (opcode == TGSI_OPCODE_TXF || opcode == TGSI_OPCODE_TXQ)
> >> +             emit_data->dst_type = LLVMVectorType(i32, 4);
> >> +     else {
> >> +             emit_data->dst_type = LLVMVectorType(
> >> +                     LLVMFloatTypeInContext(gallivm->context), 4);
> >> +
> >> +             emit_data->args[num_args++] = samp_ptr;
> >> +     }
> >> +
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm, dmask);
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm, is_rect); /* unorm */
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* r128 */
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm,
> >> +                                     tgsi_is_array_sampler(target)); /* da */
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* glc */
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* slc */
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* tfe */
> >> +     emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* lwe */
> >> +
> >> +     emit_data->arg_count = num_args;
> >> +}
> >> +
> >>  static const struct lp_build_tgsi_action tex_action;
> >>
> >>  static void tex_fetch_args(
> >> @@ -2264,6 +2322,7 @@ static void tex_fetch_args(
> >>  {
> >>       struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
> >>       struct gallivm_state *gallivm = bld_base->base.gallivm;
> >> +     LLVMBuilderRef builder = gallivm->builder;
> >>       const struct tgsi_full_instruction * inst = emit_data->inst;
> >>       unsigned opcode = inst->Instruction.Opcode;
> >>       unsigned target = inst->Texture.Texture;
> >> @@ -2278,6 +2337,8 @@ static void tex_fetch_args(
> >>       unsigned num_deriv_channels = 0;
> >>       bool has_offset = HAVE_LLVM >= 0x0305 ? inst->Texture.NumOffsets > 0 : false;
> >>       LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
> >> +     LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
> >> +     unsigned dmask = 0xf;
> >>
> >>       sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
> >>       sampler_index = emit_data->inst->Src[sampler_src].Register.Index;
> >> @@ -2308,6 +2369,43 @@ static void tex_fetch_args(
> >>               fmask_ptr = si_shader_ctx->resources[SI_FMASK_TEX_OFFSET + sampler_index];
> >>       }
> >>
> >> +     if (opcode == TGSI_OPCODE_TXQ) {
> >> +             if (target == TGSI_TEXTURE_BUFFER) {
> >> +                     LLVMTypeRef v8i32 = LLVMVectorType(i32, 8);
> >> +
> >> +                     /* Read the size from the buffer descriptor directly. */
> >> +                     LLVMValueRef res = LLVMBuildBitCast(builder, res_ptr, v8i32, "");
> >> +                     LLVMValueRef size = LLVMBuildExtractElement(builder, res,
> >> +                                                     lp_build_const_int32(gallivm, 6), "");
> >> +
> >> +                     if (si_shader_ctx->screen->b.chip_class >= VI) {
> >> +                             /* On VI, the descriptor contains the size in bytes,
> >> +                              * but TXQ must return the size in elements.
> >> +                              * The stride is always non-zero for resources using TXQ.
> >> +                              */
> >> +                             LLVMValueRef stride =
> >> +                                     LLVMBuildExtractElement(builder, res,
> >> +                                                             lp_build_const_int32(gallivm, 5), "");
> >> +                             stride = LLVMBuildLShr(builder, stride,
> >> +                                                    lp_build_const_int32(gallivm, 16), "");
> >> +                             stride = LLVMBuildAnd(builder, stride,
> >> +                                                   lp_build_const_int32(gallivm, 0x3FFF), "");
> >> +
> >> +                             size = LLVMBuildUDiv(builder, size, stride, "");
> >> +                     }
> >> +
> >> +                     emit_data->args[0] = size;
> >> +                     return;
> >> +             }
> >> +
> >> +             /* Textures - set the mip level. */
> >> +             address[count++] = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X);
> >> +
> >> +             set_tex_fetch_args(gallivm, emit_data, opcode, target, res_ptr,
> >> +                                NULL, address, count, 0xf);
> >> +             return;
> >> +     }
> >> +
> >>       if (target == TGSI_TEXTURE_BUFFER) {
> >>               LLVMTypeRef i128 = LLVMIntTypeInContext(gallivm->context, 128);
> >>               LLVMTypeRef v2i128 = LLVMVectorType(i128, 2);
> >> @@ -2457,9 +2555,7 @@ static void tex_fetch_args(
> >>
> >>       for (chan = 0; chan < count; chan++ ) {
> >>               address[chan] = LLVMBuildBitCast(gallivm->builder,
> >> -                                              address[chan],
> >> -                                              LLVMInt32TypeInContext(gallivm->context),
> >> -                                              "");
> >> +                                              address[chan], i32, "");
> >>       }
> >>
> >>       /* Adjust the sample index according to FMASK.
> >> @@ -2491,22 +2587,14 @@ static void tex_fetch_args(
> >>               }
> >>               txf_address[3] = bld_base->uint_bld.zero;
> >>
> >> -             /* Pad to a power-of-two size. */
> >> -             while (txf_count < util_next_power_of_two(txf_count))
> >> -                     txf_address[txf_count++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
> >> -
> >>               /* Read FMASK using TXF. */
> >>               inst.Instruction.Opcode = TGSI_OPCODE_TXF;
> >> -             inst.Texture.Texture = target == TGSI_TEXTURE_2D_MSAA ? TGSI_TEXTURE_2D : TGSI_TEXTURE_2D_ARRAY;
> >> +             inst.Texture.Texture = target;
> >>               txf_emit_data.inst = &inst;
> >>               txf_emit_data.chan = 0;
> >> -             txf_emit_data.dst_type = LLVMVectorType(
> >> -                     LLVMInt32TypeInContext(gallivm->context), 4);
> >> -             txf_emit_data.args[0] = lp_build_gather_values(gallivm, txf_address, txf_count);
> >> -             txf_emit_data.args[1] = fmask_ptr;
> >> -             txf_emit_data.args[2] = lp_build_const_int32(gallivm, inst.Texture.Texture);
> >> -             txf_emit_data.arg_count = 3;
> >> -
> >> +             set_tex_fetch_args(gallivm, &txf_emit_data, TGSI_OPCODE_TXF,
> >> +                                target, fmask_ptr, NULL,
> >> +                                txf_address, txf_count, 0xf);
> >>               build_tex_intrinsic(&tex_action, bld_base, &txf_emit_data);
> >>
> >>               /* Initialize some constants. */
> >> @@ -2551,9 +2639,6 @@ static void tex_fetch_args(
> >>                                       final_sample, address[sample_chan], "");
> >>       }
> >>
> >> -     /* Resource */
> >> -     emit_data->args[1] = res_ptr;
> >> -
> >>       if (opcode == TGSI_OPCODE_TXF) {
> >>               /* add tex offsets */
> >>               if (inst->Texture.NumOffsets) {
> >> @@ -2589,89 +2674,37 @@ static void tex_fetch_args(
> >>                               /* texture offsets do not apply to other texture targets */
> >>                       }
> >>               }
> >> +     }
> >>
> >> -             emit_data->args[2] = lp_build_const_int32(gallivm, target);
> >> -             emit_data->arg_count = 3;
> >> -
> >> -             emit_data->dst_type = LLVMVectorType(
> >> -                     LLVMInt32TypeInContext(gallivm->context),
> >> -                     4);
> >> -     } else if (opcode == TGSI_OPCODE_TG4 ||
> >> -                opcode == TGSI_OPCODE_LODQ ||
> >> -                has_offset) {
> >> -             unsigned is_array = target == TGSI_TEXTURE_1D_ARRAY ||
> >> -                                 target == TGSI_TEXTURE_SHADOW1D_ARRAY ||
> >> -                                 target == TGSI_TEXTURE_2D_ARRAY ||
> >> -                                 target == TGSI_TEXTURE_SHADOW2D_ARRAY ||
> >> -                                 target == TGSI_TEXTURE_CUBE_ARRAY ||
> >> -                                 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY;
> >> -             unsigned is_rect = target == TGSI_TEXTURE_RECT;
> >> -             unsigned dmask = 0xf;
> >> -
> >> -             if (opcode == TGSI_OPCODE_TG4) {
> >> -                     unsigned gather_comp = 0;
> >> -
> >> -                     /* DMASK was repurposed for GATHER4. 4 components are always
> >> -                      * returned and DMASK works like a swizzle - it selects
> >> -                      * the component to fetch. The only valid DMASK values are
> >> -                      * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
> >> -                      * (red,red,red,red) etc.) The ISA document doesn't mention
> >> -                      * this.
> >> -                      */
> >> +     if (opcode == TGSI_OPCODE_TG4) {
> >> +             unsigned gather_comp = 0;
> >>
> >> -                     /* Get the component index from src1.x for Gather4. */
> >> -                     if (!tgsi_is_shadow_sampler(target)) {
> >> -                             LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates;
> >> -                             LLVMValueRef comp_imm;
> >> -                             struct tgsi_src_register src1 = inst->Src[1].Register;
> >> +             /* DMASK was repurposed for GATHER4. 4 components are always
> >> +              * returned and DMASK works like a swizzle - it selects
> >> +              * the component to fetch. The only valid DMASK values are
> >> +              * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
> >> +              * (red,red,red,red) etc.) The ISA document doesn't mention
> >> +              * this.
> >> +              */
> >>
> >> -                             assert(src1.File == TGSI_FILE_IMMEDIATE);
> >> +             /* Get the component index from src1.x for Gather4. */
> >> +             if (!tgsi_is_shadow_sampler(target)) {
> >> +                     LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates;
> >> +                     LLVMValueRef comp_imm;
> >> +                     struct tgsi_src_register src1 = inst->Src[1].Register;
> >>
> >> -                             comp_imm = imms[src1.Index][src1.SwizzleX];
> >> -                             gather_comp = LLVMConstIntGetZExtValue(comp_imm);
> >> -                             gather_comp = CLAMP(gather_comp, 0, 3);
> >> -                     }
> >> +                     assert(src1.File == TGSI_FILE_IMMEDIATE);
> >>
> >> -                     dmask = 1 << gather_comp;
> >> +                     comp_imm = imms[src1.Index][src1.SwizzleX];
> >> +                     gather_comp = LLVMConstIntGetZExtValue(comp_imm);
> >> +                     gather_comp = CLAMP(gather_comp, 0, 3);
> >>               }
> >>
> >> -             emit_data->args[2] = samp_ptr;
> >> -             emit_data->args[3] = lp_build_const_int32(gallivm, dmask);
> >> -             emit_data->args[4] = lp_build_const_int32(gallivm, is_rect); /* unorm */
> >> -             emit_data->args[5] = lp_build_const_int32(gallivm, 0); /* r128 */
> >> -             emit_data->args[6] = lp_build_const_int32(gallivm, is_array); /* da */
> >> -             emit_data->args[7] = lp_build_const_int32(gallivm, 0); /* glc */
> >> -             emit_data->args[8] = lp_build_const_int32(gallivm, 0); /* slc */
> >> -             emit_data->args[9] = lp_build_const_int32(gallivm, 0); /* tfe */
> >> -             emit_data->args[10] = lp_build_const_int32(gallivm, 0); /* lwe */
> >> -
> >> -             emit_data->arg_count = 11;
> >> -
> >> -             emit_data->dst_type = LLVMVectorType(
> >> -                     LLVMFloatTypeInContext(gallivm->context),
> >> -                     4);
> >> -     } else {
> >> -             emit_data->args[2] = samp_ptr;
> >> -             emit_data->args[3] = lp_build_const_int32(gallivm, target);
> >> -             emit_data->arg_count = 4;
> >> -
> >> -             emit_data->dst_type = LLVMVectorType(
> >> -                     LLVMFloatTypeInContext(gallivm->context),
> >> -                     4);
> >> +             dmask = 1 << gather_comp;
> >>       }
> >>
> >> -     /* The fetch opcode has been converted to a 2D array fetch.
> >> -      * This simplifies the LLVM backend. */
> >> -     if (target == TGSI_TEXTURE_CUBE_ARRAY)
> >> -             target = TGSI_TEXTURE_2D_ARRAY;
> >> -     else if (target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
> >> -             target = TGSI_TEXTURE_SHADOW2D_ARRAY;
> >> -
> >> -     /* Pad to power of two vector */
> >> -     while (count < util_next_power_of_two(count))
> >> -             address[count++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
> >> -
> >> -     emit_data->args[0] = lp_build_gather_values(gallivm, address, count);
> >> +     set_tex_fetch_args(gallivm, emit_data, opcode, target, res_ptr,
> >> +                        samp_ptr, address, count, dmask);
> >>  }
> >>
> >>  static void build_tex_intrinsic(const struct lp_build_tgsi_action * action,
> >> @@ -2684,6 +2717,16 @@ static void build_tex_intrinsic(const struct lp_build_tgsi_action * action,
> >>       char intr_name[127];
> >>       bool has_offset = HAVE_LLVM >= 0x0305 ?
> >>                               emit_data->inst->Texture.NumOffsets > 0 : false;
> >> +     bool is_shadow = tgsi_is_shadow_sampler(target);
> >> +     char type[64];
> >> +     const char *name = "llvm.SI.image.sample";
> >> +     const char *infix = "";
> >> +
> >> +     if (opcode == TGSI_OPCODE_TXQ && target == TGSI_TEXTURE_BUFFER) {
> >> +             /* Just return the buffer size. */
> >> +             emit_data->output[emit_data->chan] = emit_data->args[0];
> >> +             return;
> >> +     }
> >>
> >>       if (target == TGSI_TEXTURE_BUFFER) {
> >>               emit_data->output[emit_data->chan] = lp_build_intrinsic(
> >> @@ -2694,191 +2737,68 @@ static void build_tex_intrinsic(const struct lp_build_tgsi_action * action,
> >>               return;
> >>       }
> >>
> >> -     if (opcode == TGSI_OPCODE_TG4 ||
> >> -         opcode == TGSI_OPCODE_LODQ ||
> >> -         (opcode != TGSI_OPCODE_TXF && has_offset)) {
> >> -             bool is_shadow = tgsi_is_shadow_sampler(target);
> >> -             const char *name = "llvm.SI.image.sample";
> >> -             const char *infix = "";
> >> -
> >> -             switch (opcode) {
> >> -             case TGSI_OPCODE_TEX:
> >> -             case TGSI_OPCODE_TEX2:
> >> -             case TGSI_OPCODE_TXP:
> >> -                     break;
> >> -             case TGSI_OPCODE_TXB:
> >> -             case TGSI_OPCODE_TXB2:
> >> -                     infix = ".b";
> >> -                     break;
> >> -             case TGSI_OPCODE_TXL:
> >> -             case TGSI_OPCODE_TXL2:
> >> -                     infix = ".l";
> >> -                     break;
> >> -             case TGSI_OPCODE_TXD:
> >> -                     infix = ".d";
> >> -                     break;
> >> -             case TGSI_OPCODE_TG4:
> >> -                     name = "llvm.SI.gather4";
> >> -                     break;
> >> -             case TGSI_OPCODE_LODQ:
> >> -                     name = "llvm.SI.getlod";
> >> -                     is_shadow = false;
> >> -                     has_offset = false;
> >> -                     break;
> >> -             default:
> >> -                     assert(0);
> >> -                     return;
> >> -             }
> >> -
> >> -             /* Add the type and suffixes .c, .o if needed. */
> >> -             sprintf(intr_name, "%s%s%s%s.v%ui32", name,
> >> -                     is_shadow ? ".c" : "", infix, has_offset ? ".o" : "",
> >> -                     LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
> >> -
> >> -             emit_data->output[emit_data->chan] = lp_build_intrinsic(
> >> -                     base->gallivm->builder, intr_name, emit_data->dst_type,
> >> -                     emit_data->args, emit_data->arg_count,
> >> -                     LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
> >> -     } else {
> >> -             LLVMTypeRef i8, v16i8, v32i8;
> >> -             const char *name;
> >> -
> >> -             switch (opcode) {
> >> -             case TGSI_OPCODE_TEX:
> >> -             case TGSI_OPCODE_TEX2:
> >> -             case TGSI_OPCODE_TXP:
> >> -                     name = "llvm.SI.sample";
> >> -                     break;
> >> -             case TGSI_OPCODE_TXB:
> >> -             case TGSI_OPCODE_TXB2:
> >> -                     name = "llvm.SI.sampleb";
> >> -                     break;
> >> -             case TGSI_OPCODE_TXD:
> >> -                     name = "llvm.SI.sampled";
> >> -                     break;
> >> -             case TGSI_OPCODE_TXF:
> >> -                     name = "llvm.SI.imageload";
> >> -                     break;
> >> -             case TGSI_OPCODE_TXL:
> >> -             case TGSI_OPCODE_TXL2:
> >> -                     name = "llvm.SI.samplel";
> >> -                     break;
> >> -             default:
> >> -                     assert(0);
> >> -                     return;
> >> -             }
> >> -
> >> -             i8 = LLVMInt8TypeInContext(base->gallivm->context);
> >> -             v16i8 = LLVMVectorType(i8, 16);
> >> -             v32i8 = LLVMVectorType(i8, 32);
> >> -
> >> -             emit_data->args[1] = LLVMBuildBitCast(base->gallivm->builder,
> >> -                                             emit_data->args[1], v32i8, "");
> >> -             if (opcode != TGSI_OPCODE_TXF) {
> >> -                     emit_data->args[2] = LLVMBuildBitCast(base->gallivm->builder,
> >> -                                             emit_data->args[2], v16i8, "");
> >> -             }
> >> -
> >> -             sprintf(intr_name, "%s.v%ui32", name,
> >> -                     LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
> >> -
> >> -             emit_data->output[emit_data->chan] = lp_build_intrinsic(
> >> -                     base->gallivm->builder, intr_name, emit_data->dst_type,
> >> -                     emit_data->args, emit_data->arg_count,
> >> -                     LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
> >> -     }
> >> -}
> >> -
> >> -static void txq_fetch_args(
> >> -     struct lp_build_tgsi_context * bld_base,
> >> -     struct lp_build_emit_data * emit_data)
> >> -{
> >> -     struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
> >> -     const struct tgsi_full_instruction *inst = emit_data->inst;
> >> -     struct gallivm_state *gallivm = bld_base->base.gallivm;
> >> -     LLVMBuilderRef builder = gallivm->builder;
> >> -     unsigned target = inst->Texture.Texture;
> >> -     LLVMValueRef res_ptr;
> >> -
> >> -     if (inst->Src[1].Register.Indirect) {
> >> -             const struct tgsi_full_src_register *reg = &inst->Src[1];
> >> -             LLVMValueRef ind_index;
> >> -
> >> -             ind_index = get_indirect_index(si_shader_ctx, &reg->Indirect, reg->Register.Index);
> >> -
> >> -             res_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_RESOURCE);
> >> -             res_ptr = build_indexed_load_const(si_shader_ctx, res_ptr,
> >> -                                                ind_index);
> >> -     } else
> >> -             res_ptr = si_shader_ctx->resources[inst->Src[1].Register.Index];
> >> -
> >> -     if (target == TGSI_TEXTURE_BUFFER) {
> >> -             LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
> >> -             LLVMTypeRef v8i32 = LLVMVectorType(i32, 8);
> >> -
> >> -             /* Read the size from the buffer descriptor directly. */
> >> -             LLVMValueRef res = LLVMBuildBitCast(builder, res_ptr, v8i32, "");
> >> -             LLVMValueRef size = LLVMBuildExtractElement(builder, res,
> >> -                                             lp_build_const_int32(gallivm, 6), "");
> >> -
> >> -             if (si_shader_ctx->screen->b.chip_class >= VI) {
> >> -                     /* On VI, the descriptor contains the size in bytes,
> >> -                      * but TXQ must return the size in elements.
> >> -                      * The stride is always non-zero for resources using TXQ.
> >> -                      */
> >> -                     LLVMValueRef stride =
> >> -                             LLVMBuildExtractElement(builder, res,
> >> -                                                     lp_build_const_int32(gallivm, 5), "");
> >> -                     stride = LLVMBuildLShr(builder, stride,
> >> -                                            lp_build_const_int32(gallivm, 16), "");
> >> -                     stride = LLVMBuildAnd(builder, stride,
> >> -                                           lp_build_const_int32(gallivm, 0x3FFF), "");
> >> -
> >> -                     size = LLVMBuildUDiv(builder, size, stride, "");
> >> -             }
> >> -
> >> -             emit_data->args[0] = size;
> >> +     switch (opcode) {
> >> +     case TGSI_OPCODE_TXF:
> >> +             name = target == TGSI_TEXTURE_2D_MSAA ||
> >> +                    target == TGSI_TEXTURE_2D_ARRAY_MSAA ?
> >> +                            "llvm.SI.image.load" :
> >> +                            "llvm.SI.image.load.mip";
> >> +             is_shadow = false;
> >> +             has_offset = false;
> >> +             break;
> >> +     case TGSI_OPCODE_TXQ:
> >> +             name = "llvm.SI.getresinfo";
> >> +             is_shadow = false;
> >> +             has_offset = false;
> >> +             break;
> >> +     case TGSI_OPCODE_LODQ:
> >> +             name = "llvm.SI.getlod";
> >> +             is_shadow = false;
> >> +             has_offset = false;
> >> +             break;
> >> +     case TGSI_OPCODE_TEX:
> >> +     case TGSI_OPCODE_TEX2:
> >> +     case TGSI_OPCODE_TXP:
> >> +             break;
> >> +     case TGSI_OPCODE_TXB:
> >> +     case TGSI_OPCODE_TXB2:
> >> +             infix = ".b";
> >> +             break;
> >> +     case TGSI_OPCODE_TXL:
> >> +     case TGSI_OPCODE_TXL2:
> >> +             infix = ".l";
> >> +             break;
> >> +     case TGSI_OPCODE_TXD:
> >> +             infix = ".d";
> >> +             break;
> >> +     case TGSI_OPCODE_TG4:
> >> +             name = "llvm.SI.gather4";
> >> +             break;
> >> +     default:
> >> +             assert(0);
> >>               return;
> >>       }
> >>
> >> -     /* Mip level */
> >> -     emit_data->args[0] = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X);
> >> -
> >> -     /* Resource */
> >> -     emit_data->args[1] = res_ptr;
> >> -
> >> -     /* Texture target */
> >> -     if (target == TGSI_TEXTURE_CUBE_ARRAY ||
> >> -         target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
> >> -             target = TGSI_TEXTURE_2D_ARRAY;
> >> -
> >> -     emit_data->args[2] = lp_build_const_int32(bld_base->base.gallivm,
> >> -                                               target);
> >> -
> >> -     emit_data->arg_count = 3;
> >> -
> >> -     emit_data->dst_type = LLVMVectorType(
> >> -             LLVMInt32TypeInContext(bld_base->base.gallivm->context),
> >> -             4);
> >> -}
> >> -
> >> -static void build_txq_intrinsic(const struct lp_build_tgsi_action * action,
> >> -                             struct lp_build_tgsi_context * bld_base,
> >> -                             struct lp_build_emit_data * emit_data)
> >> -{
> >> -     unsigned target = emit_data->inst->Texture.Texture;
> >> +     if (LLVMGetTypeKind(LLVMTypeOf(emit_data->args[0])) == LLVMVectorTypeKind)
> >> +             sprintf(type, ".v%ui32",
> >> +                     LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
> >> +     else
> >> +             strcpy(type, ".i32");
> >>
> >> -     if (target == TGSI_TEXTURE_BUFFER) {
> >> -             /* Just return the buffer size. */
> >> -             emit_data->output[emit_data->chan] = emit_data->args[0];
> >> -             return;
> >> -     }
> >> +     /* Add the type and suffixes .c, .o if needed. */
> >> +     sprintf(intr_name, "%s%s%s%s%s",
> >> +             name, is_shadow ? ".c" : "", infix,
> >> +             has_offset ? ".o" : "", type);
> >>
> >> -     build_tgsi_intrinsic_nomem(action, bld_base, emit_data);
> >> +     emit_data->output[emit_data->chan] = lp_build_intrinsic(
> >> +             base->gallivm->builder, intr_name, emit_data->dst_type,
> >> +             emit_data->args, emit_data->arg_count,
> >> +             LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
> >>
> >>       /* Divide the number of layers by 6 to get the number of cubes. */
> >> -     if (target == TGSI_TEXTURE_CUBE_ARRAY ||
> >> -         target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
> >> +     if (opcode == TGSI_OPCODE_TXQ &&
> >> +         (target == TGSI_TEXTURE_CUBE_ARRAY ||
> >> +          target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)) {
> >>               LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> >>               LLVMValueRef two = lp_build_const_int32(bld_base->base.gallivm, 2);
> >>               LLVMValueRef six = lp_build_const_int32(bld_base->base.gallivm, 6);
> >> @@ -3355,12 +3275,6 @@ static const struct lp_build_tgsi_action tex_action = {
> >>       .emit = build_tex_intrinsic,
> >>  };
> >>
> >> -static const struct lp_build_tgsi_action txq_action = {
> >> -     .fetch_args = txq_fetch_args,
> >> -     .emit = build_txq_intrinsic,
> >> -     .intr_name = "llvm.SI.resinfo"
> >> -};
> >> -
> >>  static const struct lp_build_tgsi_action interp_action = {
> >>       .fetch_args = interp_fetch_args,
> >>       .emit = build_interp_intrinsic,
> >> @@ -4070,7 +3984,7 @@ int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
> >>       bld_base->op_actions[TGSI_OPCODE_TXL] = tex_action;
> >>       bld_base->op_actions[TGSI_OPCODE_TXL2] = tex_action;
> >>       bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
> >> -     bld_base->op_actions[TGSI_OPCODE_TXQ] = txq_action;
> >> +     bld_base->op_actions[TGSI_OPCODE_TXQ] = tex_action;
> >>       bld_base->op_actions[TGSI_OPCODE_TG4] = tex_action;
> >>       bld_base->op_actions[TGSI_OPCODE_LODQ] = tex_action;
> >>
> >> --
> >> 2.1.4
> >>
> >> _______________________________________________
> >> mesa-dev mailing list
> >> mesa-dev at lists.freedesktop.org
> >> http://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list