<html> <head> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> </head> <body text="#000000" bgcolor="#FFFFFF"> <div class="moz-cite-prefix">On 6/3/19 9:33 PM, Marek Olšák wrote: </div> <blockquote type="cite" cite="mid:CAAxE2A4skc9UpFBfGTCynndCEHqjxNNbkNZznbXjuOLf_57Dew@mail.gmail.com"> <meta http-equiv="content-type" content="text/html; charset=UTF-8"> <div dir="ltr"> <div>SI doesn't support buffer_load_dwordx3 and buffer_store_dwordx3, but it supports buffer_load_format_xyz and buffer_store_format_xyz.</div> </div> </blockquote> OK, I will update. <blockquote type="cite" cite="mid:CAAxE2A4skc9UpFBfGTCynndCEHqjxNNbkNZznbXjuOLf_57Dew@mail.gmail.com"> <div dir="ltr"> <div> </div> <div>Marek </div> </div> <div class="gmail_quote"> <div dir="ltr" class="gmail_attr">On Mon, Jun 3, 2019 at 9:09 AM Samuel Pitoiset <<a href="mailto:samuel.pitoiset@gmail.com" moz-do-not-send="true">samuel.pitoiset@gmail.com</a>> wrote: </div> <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">I thought LLVM was able to handle that itself but actually it does not. That means we shouldn't try to emit vec3 on SI because it's unsupported. Fixes: 6970a9a6ca9 ("ac,radv: remove the vec3 restriction with LLVM 9+")" Signed-off-by: Samuel Pitoiset <<a href="mailto:samuel.pitoiset@gmail.com" target="_blank" moz-do-not-send="true">samuel.pitoiset@gmail.com</a>> --- src/amd/common/ac_llvm_build.c | 12 ++++++------ src/amd/common/ac_llvm_util.h | 9 +++++++++ src/amd/common/ac_nir_to_llvm.c | 3 ++- src/amd/vulkan/radv_nir_to_llvm.c | 2 +- 4 files changed, 18 insertions(+), 8 deletions(-) diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c index 613c1eef942..7f5c8ef873c 100644 --- a/src/amd/common/ac_llvm_build.c +++ b/src/amd/common/ac_llvm_build.c @@ -1167,7 +1167,7 @@ ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx, args[idx++] = voffset ? voffset : ctx->i32_0; args[idx++] = soffset ? soffset : ctx->i32_0; args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0); - unsigned func = HAVE_LLVM < 0x900 && num_channels == 3 ? 4 : num_channels; + unsigned func = !ac_has_vec3_support(ctx->chip_class) && num_channels == 3 ? 4 : num_channels; const char *indexing_kind = structurized ? "struct" : "raw"; char name[256], type_name[8]; @@ -1227,7 +1227,7 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx, { /* Split 3 channel stores, because only LLVM 9+ support 3-channel * intrinsics. */ - if (num_channels == 3 && HAVE_LLVM < 0x900) { + if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class)) { LLVMValueRef v[3], v01; for (int i = 0; i < 3; i++) { @@ -1354,7 +1354,7 @@ ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx, args[idx++] = voffset ? voffset : ctx->i32_0; args[idx++] = soffset ? soffset : ctx->i32_0; args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0); - unsigned func = HAVE_LLVM < 0x900 && num_channels == 3 ? 4 : num_channels; + unsigned func = !ac_has_vec3_support(ctx->chip_class) && num_channels == 3 ? 4 : num_channels; const char *indexing_kind = structurized ? "struct" : "raw"; char name[256], type_name[8]; @@ -1420,7 +1420,7 @@ ac_build_buffer_load(struct ac_llvm_context *ctx, if (num_channels == 1) return result[0]; - if (num_channels == 3 && HAVE_LLVM < 0x900) + if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class)) result[num_channels++] = LLVMGetUndef(ctx->f32); return ac_build_gather_values(ctx, result, num_channels); } @@ -1512,7 +1512,7 @@ ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx, args[idx++] = soffset ? soffset : ctx->i32_0; args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0); args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0); - unsigned func = HAVE_LLVM < 0x900 && num_channels == 3 ? 4 : num_channels; + unsigned func = !ac_has_vec3_support(ctx->chip_class) && num_channels == 3 ? 4 : num_channels; const char *indexing_kind = structurized ? "struct" : "raw"; char name[256], type_name[8]; @@ -2011,7 +2011,7 @@ ac_build_llvm8_tbuffer_store(struct ac_llvm_context *ctx, args[idx++] = soffset ? soffset : ctx->i32_0; args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0); args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0); - unsigned func = HAVE_LLVM < 0x900 && num_channels == 3 ? 4 : num_channels; + unsigned func = !ac_has_vec3_support(ctx->chip_class) && num_channels == 3 ? 4 : num_channels; const char *indexing_kind = structurized ? "struct" : "raw"; char name[256], type_name[8]; diff --git a/src/amd/common/ac_llvm_util.h b/src/amd/common/ac_llvm_util.h index ca00540da80..a45647a3360 100644 --- a/src/amd/common/ac_llvm_util.h +++ b/src/amd/common/ac_llvm_util.h @@ -146,6 +146,15 @@ bool ac_compile_module_to_binary(struct ac_compiler_passes *p, LLVMModuleRef mod void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr); void ac_enable_global_isel(LLVMTargetMachineRef tm); +static inline bool +ac_has_vec3_support(enum chip_class chip) +{ + if (chip == GFX6) + return false; + + return HAVE_LLVM >= 0x900; +} + #ifdef __cplusplus } #endif diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c index 51f92a6b062..429dac63d63 100644 --- a/src/amd/common/ac_nir_to_llvm.c +++ b/src/amd/common/ac_nir_to_llvm.c @@ -1576,7 +1576,8 @@ static void visit_store_ssbo(struct ac_nir_context *ctx, /* Due to an LLVM limitation with LLVM < 9, split 3-element * writes into a 2-element and a 1-element write. */ - if (count == 3 && (elem_size_bytes != 4 || HAVE_LLVM < 0x900)) { + if (count == 3 && + (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class))) { writemask |= 1 << (start + 2); count = 2; } diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c index dca4bebcdd1..ab552ae34ab 100644 --- a/src/amd/vulkan/radv_nir_to_llvm.c +++ b/src/amd/vulkan/radv_nir_to_llvm.c @@ -2768,7 +2768,7 @@ radv_emit_stream_output(struct radv_shader_context *ctx, /* fall through */ case 4: /* as v4i32 */ vdata = ac_build_gather_values(&ctx->ac, out, - HAVE_LLVM < 0x900 ? + !ac_has_vec3_support(ctx->ac.chip_class) ? util_next_power_of_two(num_comps) : num_comps); break; -- 2.21.0 _______________________________________________ mesa-dev mailing list <a href="mailto:mesa-dev@lists.freedesktop.org" target="_blank" moz-do-not-send="true">mesa-dev@lists.freedesktop.org</a> <a href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev" rel="noreferrer" target="_blank" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a></blockquote> </div> </blockquote> </body> </html>