<div dir="auto"><div><div class="gmail_extra"><div class="gmail_quote">On Jan 4, 2017 5:46 AM, "Juan A. Suarez Romero" <<a href="mailto:jasuarez@igalia.com">jasuarez@igalia.com</a>> wrote:<br type="attribution"><blockquote class="quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div class="quoted-text"><div>On Tue, 2017-01-03 at 14:41 -0800, Jason Ekstrand wrote:</div></div><div class="elided-text"><blockquote type="cite"><div dir="ltr"><div dir="auto"><div><div class="gmail_extra"><div class="gmail_quote">I made a few pretty trivial comments. With those addressed,<br><br></div><div class="gmail_quote">Reviewed-by: Jason Ekstrand <<a href="mailto:jason@jlekstrand.net" target="_blank">jason@jlekstrand.net</a>><br></div><div class="gmail_quote" dir="auto"><br></div><div class="gmail_quote">On Dec 16, 2016 8:55 AM, "Juan A. Suarez Romero" <<a href="mailto:jasuarez@igalia.com" target="_blank">jasuarez@igalia.com</a>> wrote:<br type="attribution"><blockquote class="m_8367481713989651032m_6457694654742999625quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">So far, input_reads was a bitmap tracking which vertex input locations<br> were being used.<br> <br> In OpenGL, an attribute bigger than a vec4 (like a dvec3 or dvec4)<br> consumes just one location, any other small attribute. So we mark the<br> proper bit in inputs_read, and also the same bit in double_inputs_read<br> if the attribute is a dvec3/dvec4.<br> <br> But in Vulkan, this is slightly different: a dvec3/dvec4 attribute<br> consumes two locations, not just one. And hence two bits would be marked<br> in inputs_read for the same vertex input attribute.<br> <br> To avoid handling two different situations in NIR, we just choose the<br> latest one: in OpenGL, when creating NIR from GLSL/IR, any dvec3/dvec4<br> vertex input attribute is marked with two bits in the inputs_read bitmap<br> (and also in the double_inputs_read), and following attributes are<br> adjusted accordingly.<br> <br> As example, if in our GLSL/IR shader we have three attributes:<br> <br> layout(location = 0) vec3 attr0;<br> layout(location = 1) dvec4 attr1;<br> layout(location = 2) dvec3 attr2;<br> <br> then in our NIR shader we put attr0 in location 0, attr1 in locations 1<br> and 2, and attr2 in location 3.<br></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">attr2 goes in locations 3 *and* 4, correct?</div><div dir="auto"><br></div><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="m_8367481713989651032m_6457694654742999625quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> Checking carefully, basically we are using slots rather than locations<br> in NIR.<br> <br> When emitting the vertices, we do a inverse map to know the<br> corresponding location for each slot.<br> <br> v2 (Jason):<br> - use two slots from inputs_read for dvec3/dvec4 NIR from GLSL/IR.<br> ---<br> src/compiler/glsl/glsl_to_nir<wbr>.cpp | 28 +++++++++++++<br> src/compiler/nir/nir_gather_i<wbr>nfo.c | 48 ++++++++++-----------<br> src/intel/vulkan/genX_pipelin<wbr>e.c | 63 +++++++++++++++++-----------<br> src/mesa/drivers/dri/i965/brw<wbr>_draw_upload.c | 11 +++--<br> src/mesa/drivers/dri/i965/brw<wbr>_fs.cpp | 13 ------<br> src/mesa/drivers/dri/i965/brw<wbr>_fs_visitor.cpp | 3 +-<br> src/mesa/drivers/dri/i965/brw<wbr>_nir.c | 6 +--<br> src/mesa/drivers/dri/i965/brw<wbr>_nir.h | 1 -<br> src/mesa/drivers/dri/i965/brw<wbr>_vec4.cpp | 11 +++--<br> 9 files changed, 106 insertions(+), 78 deletions(-)<br> <br> diff --git a/src/compiler/glsl/glsl_to_ni<wbr>r.cpp b/src/compiler/glsl/glsl_to_ni<wbr>r.cpp<br> index 4debc37..0814dad 100644<br> --- a/src/compiler/glsl/glsl_to_ni<wbr>r.cpp<br> +++ b/src/compiler/glsl/glsl_to_ni<wbr>r.cpp<br> @@ -129,6 +129,19 @@ private:<br> <br> } /* end of anonymous namespace */<br> <br> +static void<br> +nir_remap_attributes(nir_shad<wbr>er *shader)<br> +{<br> + nir_foreach_variable(var, &shader->inputs) {<br> + var->data.location += _mesa_bitcount_64(shader->info<wbr>->double_inputs_read &<br> + BITFIELD64_MASK(var->data.loca<wbr>tion));<br> + }<br> +<br> + /* Once the remap is done, reset double_inputs_read, so later it will have<br> + * which location/slots are doubles */<br> + shader->info->double_inputs_r<wbr>ead = 0;<br> +}<br> +<br> nir_shader *<br> glsl_to_nir(const struct gl_shader_program *shader_prog,<br> gl_shader_stage stage,<br> @@ -146,6 +159,13 @@ glsl_to_nir(const struct gl_shader_program *shader_prog,<br> <br> nir_lower_constant_initializer<wbr>s(shader, (nir_variable_mode)~0);<br> <br> + /* Remap the locations to slots so those requiring two slots will occupy<br> + * two locations. For instance, if we have in the IR code a dvec3 attr0 in<br> + * location 0 and vec4 attr1 in location 1, in NIR attr0 will use<br> + * locations/slots 0 and 1, and attr1 will use location/slot 2 */<br> + if (shader->stage == MESA_SHADER_VERTEX)<br> + nir_remap_attributes(shader);<br> +<br> shader->info->name = ralloc_asprintf(shader, "GLSL%d", shader_prog->Name);<br> if (shader_prog->Label)<br> shader->info->label = ralloc_strdup(shader, shader_prog->Label);<br> @@ -315,6 +335,14 @@ nir_visitor::visit(ir_variable *ir)<br> } else {<br> var->data.mode = nir_var_shader_in;<br> }<br> +<br> + /* Mark all the locations that require two slots */<br> + if (glsl_type_is_dual_slot(glsl_w<wbr>ithout_array(var->type))) {<br> + for (uint i = 0; i < glsl_count_attribute_slots(var<wbr>->type, true); i++) {<br> + uint64_t bitfield = BITFIELD64_BIT(var->data.locat<wbr>ion + i);<br> + shader->info->double_inputs_re<wbr>ad |= bitfield;<br> + }<br> + }<br> break;<br> <br> case ir_var_shader_out:<br> diff --git a/src/compiler/nir/nir_gather_<wbr>info.c b/src/compiler/nir/nir_gather_<wbr>info.c<br> index 07c9949..35a1ce4 100644<br> --- a/src/compiler/nir/nir_gather_<wbr>info.c<br> +++ b/src/compiler/nir/nir_gather_<wbr>info.c<br> @@ -53,11 +53,6 @@ set_io_mask(nir_shader *shader, nir_variable *var, int offset, int len)<br> else<br> shader->info->inputs_read |= bitfield;<br> <br> - /* double inputs read is only for vertex inputs */<br> - if (shader->stage == MESA_SHADER_VERTEX &&<br> - glsl_type_is_dual_slot(glsl_w<wbr>ithout_array(var->type)))<br> - shader->info->double_inputs_re<wbr>ad |= bitfield;<br> -<br> if (shader->stage == MESA_SHADER_FRAGMENT) {<br> shader->info->fs.uses_sample_<wbr>qualifier |= var->data.sample;<br> }<br> @@ -83,26 +78,21 @@ static void<br> mark_whole_variable(nir_shade<wbr>r *shader, nir_variable *var)<br> {<br> const struct glsl_type *type = var->type;<br> - bool is_vertex_input = false;<br> <br> if (nir_is_per_vertex_io(var, shader->stage)) {<br> assert(glsl_type_is_array(typ<wbr>e));<br> type = glsl_get_array_element(type);<br> }<br> <br> - if (shader->stage == MESA_SHADER_VERTEX &&<br> - var->data.mode == nir_var_shader_in)<br> - is_vertex_input = true;<br> -<br> const unsigned slots =<br> var->data.compact ? DIV_ROUND_UP(glsl_get_length(t<wbr>ype), 4)<br> - : glsl_count_attribute_slots(typ<wbr>e, is_vertex_input);<br> + : glsl_count_attribute_slots(typ<wbr>e, false);<br> <br> set_io_mask(shader, var, 0, slots);<br> }<br> <br> static unsigned<br> -get_io_offset(nir_deref_var *deref, bool is_vertex_input)<br> +get_io_offset(nir_deref_var *deref)<br> {<br> unsigned offset = 0;<br> <br> @@ -117,7 +107,7 @@ get_io_offset(nir_deref_var *deref, bool is_vertex_input)<br> return -1;<br> }<br> <br> - offset += glsl_count_attribute_slots(tai<wbr>l->type, is_vertex_input) *<br> + offset += glsl_count_attribute_slots(tai<wbr>l->type, false) *<br> deref_array->base_offset;<br> }<br> /* TODO: we can get the offset for structs here see nir_lower_io() */<br> @@ -163,12 +153,7 @@ try_mask_partial_io(nir_shader *shader, nir_deref_var *deref)<br> return false;<br> }<br> <br> - bool is_vertex_input = false;<br> - if (shader->stage == MESA_SHADER_VERTEX &&<br> - var->data.mode == nir_var_shader_in)<br> - is_vertex_input = true;<br> -<br> - unsigned offset = get_io_offset(deref, is_vertex_input);<br> + unsigned offset = get_io_offset(deref);<br> if (offset == -1)<br> return false;<br> <br> @@ -184,8 +169,7 @@ try_mask_partial_io(nir_shader *shader, nir_deref_var *deref)<br> }<br> <br> /* double element width for double types that takes two slots */<br> - if (!is_vertex_input &&<br> - glsl_type_is_dual_slot(glsl_w<wbr>ithout_array(type))) {<br> + if (glsl_type_is_dual_slot(glsl_w<wbr>ithout_array(type))) {<br> elem_width *= 2;<br> }<br> <br> @@ -220,13 +204,27 @@ gather_intrinsic_info(nir_intr<wbr>insic_instr *instr, nir_shader *shader)<br> case nir_intrinsic_interp_var_at_sa<wbr>mple:<br> case nir_intrinsic_interp_var_at_of<wbr>fset:<br> case nir_intrinsic_load_var:<br> - case nir_intrinsic_store_var:<br> - if (instr->variables[0]->var->dat<wbr>a.mode == nir_var_shader_in ||<br> - instr->variables[0]->var->data<wbr>.mode == nir_var_shader_out) {<br> + case nir_intrinsic_store_var: {<br> + nir_variable *var = instr->variables[0]->var;<br> +<br> + if (var->data.mode == nir_var_shader_in ||<br> + var->data.mode == nir_var_shader_out) {<br> if (!try_mask_partial_io(shader, instr->variables[0]))<br> - mark_whole_variable(shader, instr->variables[0]->var);<br> + mark_whole_variable(shader, var);<br> +<br> + /* We need to track which input_reads bits correspond to a<br> + * dvec3/dvec4 input attribute */<br> + if (shader->stage == MESA_SHADER_VERTEX &&<br> + var->data.mode == nir_var_shader_in &&<br> + glsl_type_is_dual_slot(glsl_w<wbr>ithout_array(var->type))) {<br> + for (uint i = 0; i < glsl_count_attribute_slots(var<wbr>->type, false); i++) {<br> + int idx = var->data.location + i;<br> + shader->info->double_inputs_r<wbr>ead |= BITFIELD64_BIT(idx);<br> + }<br> + }<br> }<br> break;<br> + }<br> <br> case nir_intrinsic_load_draw_id:<br> case nir_intrinsic_load_front_face:<br> diff --git a/src/intel/vulkan/genX_pipeli<wbr>ne.c b/src/intel/vulkan/genX_pipeli<wbr>ne.c<br> index 845d020..7b94959 100644<br> --- a/src/intel/vulkan/genX_pipeli<wbr>ne.c<br> +++ b/src/intel/vulkan/genX_pipeli<wbr>ne.c<br> @@ -33,26 +33,33 @@<br> static uint32_t<br> vertex_element_comp_control(e<wbr>num isl_format format, unsigned comp)<br> {<br> - uint8_t bits;<br> switch (comp) {<br> - case 0: bits = isl_format_layouts[format].cha<wbr>nnels.r.bits; break;<br> - case 1: bits = isl_format_layouts[format].cha<wbr>nnels.g.bits; break;<br> - case 2: bits = isl_format_layouts[format].cha<wbr>nnels.b.bits; break;<br> - case 3: bits = isl_format_layouts[format].cha<wbr>nnels.a.bits; break;<br> - default: unreachable("Invalid component");<br> - }<br> -<br> - if (bits) {<br> - return VFCOMP_STORE_SRC;<br> - } else if (comp < 3) {<br> - return VFCOMP_STORE_0;<br> - } else if (isl_format_layouts[format].ch<wbr>annels.r.type == ISL_UINT ||<br> - isl_format_layouts[format].cha<wbr>nnels.r.type == ISL_SINT) {<br> - assert(comp == 3);<br> - return VFCOMP_STORE_1_INT;<br> - } else {<br> - assert(comp == 3);<br> - return VFCOMP_STORE_1_FP;<br> + case 0:<br> + return isl_format_layouts[format].cha<wbr>nnels.r.bits ?<br> + VFCOMP_STORE_SRC : VFCOMP_STORE_0;<br> + case 1:<br> + return isl_format_layouts[format].cha<wbr>nnels.g.bits ?<br> + VFCOMP_STORE_SRC : VFCOMP_STORE_0;<br> + case 2:<br> + return isl_format_layouts[format].cha<wbr>nnels.b.bits ?<br> + VFCOMP_STORE_SRC : ((isl_format_layouts[format].c<wbr>hannels.r.type == ISL_RAW) ?<br> + VFCOMP_NOSTORE : VFCOMP_STORE_0);<br></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">Given all the line wrapping, I think it would be clearer to just have an if ladder here</div><div dir="auto"><br></div><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="m_8367481713989651032m_6457694654742999625quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + case 3:<br> + if (isl_format_layouts[format].ch<wbr>annels.a.bits)<br> + return VFCOMP_STORE_SRC;<br> + else<br></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">Please use braces when one side of the if/else is multiple lines.</div><div dir="auto"><br></div><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="m_8367481713989651032m_6457694654742999625quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + switch (isl_format_layouts[format].ch<wbr>annels.r.type) {<br> + case ISL_RAW:<br> + return isl_format_layouts[format].cha<wbr>nnels.b.bits ?<br> + VFCOMP_STORE_0 : VFCOMP_NOSTORE;<br></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">This seems a bit odd. Mind explaining what's going on here?</div><div dir="auto"><br></div></div></div></blockquote><div><br></div></div><div>Yes. When emitting 64-bit components, we either write 128 or 256 bits, using VFCOMP_STORE_0 to pad output properly. We write 256 bits if we need to emit Blue and/or Alpha components.</div><div><br></div><div>If we only need to write 128bits then we use VFCOMP_NOSTORE for the 3rd and 4th components.</div><div><br></div><div>In above code, we already know we are not emitting Alpha (the first condition in "case 3" is false). If we neither need to emit Blue component, then we only need to write 128 bits, so we return NOSTORE.</div><div><br></div><div>But if Blue is required, then we need to write 256bits, so we return VFCOMP_STORE_0 to pad the output.</div></div></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">Maybe this would be clearer: leave it structured the way it was with the "bits" temporary and add a new case to the if ladder right after the first one that is</div><div dir="auto"><br></div><div dir="auto">} else if (comp >= 2 && !<span style="font-family:sans-serif">isl_format_layouts[format].c</span><wbr style="font-family:sans-serif"><span style="font-family:sans-serif">hannels.b.bits &&</span></div><div dir="auto"><span style="font-family:sans-serif"> isl_format_layouts[format].c</span><wbr style="font-family:sans-serif"><span style="font-family:sans-serif">hannels.r.type == ISL_RAW)</span><span style="font-family:sans-serif"> {</span></div><div dir="auto"><span style="font-family:sans-serif"> /* comment about writing 128-bit chunks */</span></div><div dir="auto"><span style="font-family:sans-serif"> return VFCOMP_NOSTORE;</span></div><div dir="auto"><br></div><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div class="elided-text"><blockquote type="cite"><div dir="ltr"><div dir="auto"><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="m_8367481713989651032m_6457694654742999625quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + case ISL_UINT:<br> + case ISL_SINT:<br> + return VFCOMP_STORE_1_INT;<br> + default:<br> + return VFCOMP_STORE_1_FP;<br> + }<br> + default:<br> + unreachable("Invalid component");<br> }<br> }<br> <br> @@ -64,8 +71,10 @@ emit_vertex_input(struct anv_pipeline *pipeline,<br> <br> /* Pull inputs_read out of the VS prog data */<br> const uint64_t inputs_read = vs_prog_data->inputs_read;<br> + const uint64_t double_inputs_read = vs_prog_data->double_inputs_re<wbr>ad;<br> assert((inputs_read & ((1 << VERT_ATTRIB_GENERIC0) - 1)) == 0);<br> const uint32_t elements = inputs_read >> VERT_ATTRIB_GENERIC0;<br> + const uint32_t elements_double = double_inputs_read >> VERT_ATTRIB_GENERIC0;<br> <br> #if GEN_GEN >= 8<br> /* On BDW+, we only need to allocate space for base ids. Setting up<br> @@ -83,13 +92,16 @@ emit_vertex_input(struct anv_pipeline *pipeline,<br> vs_prog_data->uses_baseinstan<wbr>ce;<br> #endif<br> <br> - uint32_t elem_count = __builtin_popcount(elements) + needs_svgs_elem;<br> - if (elem_count == 0)<br> + uint32_t elem_count = __builtin_popcount(elements) -<br> + DIV_ROUND_UP(__builtin_popcoun<wbr>t(elements_double), 2);<br></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">This had better be divisible by 2...</div><div dir="auto"><br></div></div></div></blockquote><div><br></div></div><div>Yes, it is divisible by 2. Totally forgot to remove the DIV_ROUND_UP() in the series. We'll fix it. Thanks for the catch up!</div><div class="elided-text"><div><br></div><blockquote type="cite"><div dir="ltr"><div dir="auto"><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="m_8367481713989651032m_6457694654742999625quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> +<br> + uint32_t total_elems = elem_count + needs_svgs_elem;<br> + if (total_elems == 0)<br> return;<br> <br> uint32_t *p;<br> <br> - const uint32_t num_dwords = 1 + elem_count * 2;<br> + const uint32_t num_dwords = 1 + total_elems * 2;<br> p = anv_batch_emitn(&pipeline->bat<wbr>ch, num_dwords,<br> GENX(3DSTATE_VERTEX_ELEMENTS))<wbr>;<br> memset(p + 1, 0, (num_dwords - 1) * 4);<br> @@ -107,7 +119,10 @@ emit_vertex_input(struct anv_pipeline *pipeline,<br> if ((elements & (1 << desc->location)) == 0)<br> continue; /* Binding unused */<br> <br> - uint32_t slot = __builtin_popcount(elements & ((1 << desc->location) - 1));<br> + uint32_t slot =<br> + __builtin_popcount(elements & ((1 << desc->location) - 1)) -<br> + DIV_ROUND_UP(__builtin_popcou<wbr>nt(elements_double &<br> + ((1 << desc->location) -1)), 2);<br> <br> struct GENX(VERTEX_ELEMENT_STATE) element = {<br> .VertexBufferIndex = desc->binding,<br> @@ -137,7 +152,7 @@ emit_vertex_input(struct anv_pipeline *pipeline,<br> #endif<br> }<br> <br> - const uint32_t id_slot = __builtin_popcount(elements);<br> + const uint32_t id_slot = elem_count;<br> if (needs_svgs_elem) {<br> /* From the Broadwell PRM for the 3D_Vertex_Component_Control enum:<br> * "Within a VERTEX_ELEMENT_STATE structure, if a Component<br> diff --git a/src/mesa/drivers/dri/i965/br<wbr>w_draw_upload.c b/src/mesa/drivers/dri/i965/br<wbr>w_draw_upload.c<br> index b138cb7..c3655d8 100644<br> --- a/src/mesa/drivers/dri/i965/br<wbr>w_draw_upload.c<br> +++ b/src/mesa/drivers/dri/i965/br<wbr>w_draw_upload.c<br> @@ -472,11 +472,16 @@ brw_prepare_vertices(struct brw_context *brw)<br> /* Accumulate the list of enabled arrays. */<br> brw->vb.nr_enabled = 0;<br> while (vs_inputs) {<br> - GLuint index = ffsll(vs_inputs) - 1;<br> + GLuint first = ffsll(vs_inputs) - 1;<br> + GLuint index =<br> + first - DIV_ROUND_UP(_mesa_bitcount_64<wbr>(vs_prog_data->double_inputs_r<wbr>ead &<br> + BITFIELD64_MASK(first)), 2);<br> struct brw_vertex_element *input = &brw->vb.inputs[index];<br> input->is_dual_slot = brw->gen >= 8 &&<br> - (vs_prog_data->double_inputs_<wbr>read & BITFIELD64_BIT(index)) != 0;<br> - vs_inputs &= ~BITFIELD64_BIT(index);<br> + (vs_prog_data->double_inputs_<wbr>read & BITFIELD64_BIT(first)) != 0;<br> + vs_inputs &= ~BITFIELD64_BIT(first);<br> + if (input->is_dual_slot)<br> + vs_inputs &= ~BITFIELD64_BIT(first + 1);<br> brw->vb.enabled[brw->vb.nr_en<wbr>abled++] = input;<br> }<br> <br> diff --git a/src/mesa/drivers/dri/i965/br<wbr>w_fs.cpp b/src/mesa/drivers/dri/i965/br<wbr>w_fs.cpp<br> index c218f56..b383d98 100644<br> --- a/src/mesa/drivers/dri/i965/br<wbr>w_fs.cpp<br> +++ b/src/mesa/drivers/dri/i965/br<wbr>w_fs.cpp<br> @@ -492,19 +492,6 @@ type_size_scalar(const struct glsl_type *type)<br> return 0;<br> }<br> <br> -/* Attribute arrays are loaded as one vec4 per element (or matrix column),<br> - * except for double-precision types, which are loaded as one dvec4.<br> - */<br> -extern "C" int<br> -type_size_vs_input(const struct glsl_type *type)<br> -{<br> - if (type->is_double()) {<br> - return type_size_dvec4(type);<br> - } else {<br> - return type_size_vec4(type);<br> - }<br> -}<br> -<br> /**<br> * Create a MOV to read the timestamp register.<br> *<br> diff --git a/src/mesa/drivers/dri/i965/br<wbr>w_fs_visitor.cpp b/src/mesa/drivers/dri/i965/br<wbr>w_fs_visitor.cpp<br> index 14415bd..5c356d6 100644<br> --- a/src/mesa/drivers/dri/i965/br<wbr>w_fs_visitor.cpp<br> +++ b/src/mesa/drivers/dri/i965/br<wbr>w_fs_visitor.cpp<br> @@ -36,8 +36,7 @@ fs_reg *<br> fs_visitor::emit_vs_system_va<wbr>lue(int location)<br> {<br> fs_reg *reg = new(this->mem_ctx)<br> - fs_reg(ATTR, 4 * (_mesa_bitcount_64(nir->info-><wbr>inputs_read) +<br> - _mesa_bitcount_64(nir->info->d<wbr>ouble_inputs_read)),<br> + fs_reg(ATTR, 4 * _mesa_bitcount_64(nir->info->i<wbr>nputs_read),<br> BRW_REGISTER_TYPE_D);<br> struct brw_vs_prog_data *vs_prog_data = brw_vs_prog_data(prog_data);<br> <br> diff --git a/src/mesa/drivers/dri/i965/br<wbr>w_nir.c b/src/mesa/drivers/dri/i965/br<wbr>w_nir.c<br> index 763e3ec..24334c8 100644<br> --- a/src/mesa/drivers/dri/i965/br<wbr>w_nir.c<br> +++ b/src/mesa/drivers/dri/i965/br<wbr>w_nir.c<br> @@ -113,9 +113,7 @@ remap_vs_attrs(nir_block *block, shader_info *nir_info)<br> int attr = intrin->const_index[0];<br> int slot = _mesa_bitcount_64(nir_info->in<wbr>puts_read &<br> BITFIELD64_MASK(attr));<br> - int dslot = _mesa_bitcount_64(nir_info->do<wbr>uble_inputs_read &<br> - BITFIELD64_MASK(attr));<br> - intrin->const_index[0] = 4 * (slot + dslot);<br> + intrin->const_index[0] = 4 * slot;<br> }<br> }<br> return true;<br> @@ -204,7 +202,7 @@ brw_nir_lower_vs_inputs(nir_sh<wbr>ader *nir,<br> * loaded as one vec4 or dvec4 per element (or matrix column), depending on<br> * whether it is a double-precision type or not.<br> */<br> - nir_lower_io(nir, nir_var_shader_in, type_size_vs_input, 0);<br> + nir_lower_io(nir, nir_var_shader_in, type_size_vec4, 0);<br> <br> /* This pass needs actual constants */<br> nir_opt_constant_folding(nir);<br> diff --git a/src/mesa/drivers/dri/i965/br<wbr>w_nir.h b/src/mesa/drivers/dri/i965/br<wbr>w_nir.h<br> index 3c774d0..04c7ef8 100644<br> --- a/src/mesa/drivers/dri/i965/br<wbr>w_nir.h<br> +++ b/src/mesa/drivers/dri/i965/br<wbr>w_nir.h<br> @@ -34,7 +34,6 @@ extern "C" {<br> int type_size_scalar(const struct glsl_type *type);<br> int type_size_vec4(const struct glsl_type *type);<br> int type_size_dvec4(const struct glsl_type *type);<br> -int type_size_vs_input(const struct glsl_type *type);<br> <br> static inline int<br> type_size_scalar_bytes(const struct glsl_type *type)<br> diff --git a/src/mesa/drivers/dri/i965/br<wbr>w_vec4.cpp b/src/mesa/drivers/dri/i965/br<wbr>w_vec4.cpp<br> index b9e592f..852e88f 100644<br> --- a/src/mesa/drivers/dri/i965/br<wbr>w_vec4.cpp<br> +++ b/src/mesa/drivers/dri/i965/br<wbr>w_vec4.cpp<br> @@ -2127,7 +2127,7 @@ brw_compile_vs(const struct brw_compiler *compiler, void *log_data,<br> ((1 << shader->info->cull_distance_ar<wbr>ray_size) - 1) <<<br> shader->info->clip_distance_a<wbr>rray_size;<br> <br> - unsigned nr_attributes = _mesa_bitcount_64(prog_data->i<wbr>nputs_read);<br> + unsigned nr_attribute_slots = _mesa_bitcount_64(prog_data->i<wbr>nputs_read);<br> <br> /* gl_VertexID and gl_InstanceID are system values, but arrive via an<br> * incoming vertex attribute. So, add an extra slot.<br> @@ -2137,18 +2137,17 @@ brw_compile_vs(const struct brw_compiler *compiler, void *log_data,<br> BITFIELD64_BIT(SYSTEM_VALUE_B<wbr>ASE_INSTANCE) |<br> BITFIELD64_BIT(SYSTEM_VALUE_V<wbr>ERTEX_ID_ZERO_BASE) |<br> BITFIELD64_BIT(SYSTEM_VALUE_I<wbr>NSTANCE_ID))) {<br> - nr_attributes++;<br> + nr_attribute_slots++;<br> }<br> <br> /* gl_DrawID has its very own vec4 */<br> if (shader->info->system_values_r<wbr>ead &<br> BITFIELD64_BIT(SYSTEM_VALUE_DR<wbr>AW_ID)) {<br> - nr_attributes++;<br> + nr_attribute_slots++;<br> }<br> <br> - unsigned nr_attribute_slots =<br> - nr_attributes +<br> - _mesa_bitcount_64(shader->info<wbr>->double_inputs_read);<br> + unsigned nr_attributes = nr_attribute_slots -<br> + DIV_ROUND_UP(_mesa_bitcount_64<wbr>(shader->info->double_inputs_r<wbr>ead), 2);<br> <br> /* The 3DSTATE_VS documentation lists the lower bound on "Vertex URB Entry<br> * Read Length" as 1 in vec4 mode, and 0 in SIMD8 mode. Empirically, in<br></blockquote></div></div></div></div> </div> </blockquote></div></div></blockquote></div><br></div></div></div>