/**************************************************************************
*
* Copyright 2009 VMware, Inc.
- * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * Copyright 2007 VMware, Inc.
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* 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 TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * IN NO EVENT SHALL VMWARE AND/OR ITS 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 "util/u_format.h"
#include "util/u_dump.h"
#include "util/u_string.h"
-#include "util/u_simple_list.h"
+#include "util/simple_list.h"
#include "util/u_dual_blend.h"
-#include "os/os_time.h"
+#include "util/os_time.h"
#include "pipe/p_shader_tokens.h"
#include "draw/draw_context.h"
#include "tgsi/tgsi_dump.h"
#include "gallivm/lp_bld_flow.h"
#include "gallivm/lp_bld_debug.h"
#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_bitarit.h"
#include "gallivm/lp_bld_pack.h"
#include "gallivm/lp_bld_format.h"
#include "gallivm/lp_bld_quad.h"
#include "lp_tex_sample.h"
#include "lp_flush.h"
#include "lp_state_fs.h"
+#include "lp_rast.h"
/** Fragment shader number (for debugging) */
struct lp_type mask_type;
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef bits[16];
- LLVMValueRef mask;
+ LLVMValueRef mask, bits_vec;
int shift, i;
/*
* XXX: We'll need a different path for 16 x u8
*/
assert(fs_type.width == 32);
- assert(fs_type.length <= Elements(bits));
+ assert(fs_type.length <= ARRAY_SIZE(bits));
mask_type = lp_int_type(fs_type);
/*
for (i = 0; i < fs_type.length / 4; i++) {
unsigned j = 2 * (i % 2) + (i / 2) * 8;
- bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
- bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
- bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
- bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
+ bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
+ bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
+ bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
+ bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
}
- mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
+ bits_vec = LLVMConstVector(bits, fs_type.length);
+ mask = LLVMBuildAnd(builder, mask, bits_vec, "");
/*
- * mask = mask != 0 ? ~0 : 0
+ * mask = mask == bits ? ~0 : 0
*/
mask = lp_build_compare(gallivm,
- mask_type, PIPE_FUNC_NOTEQUAL,
- mask,
- lp_build_const_int_vec(gallivm, mask_type, 0));
+ mask_type, PIPE_FUNC_EQUAL,
+ mask, bits_vec);
return mask;
}
}
+/**
+ * Fetch the specified lp_jit_viewport structure for a given viewport_index.
+ */
+static LLVMValueRef
+lp_llvm_viewport(LLVMValueRef context_ptr,
+ struct gallivm_state *gallivm,
+ LLVMValueRef viewport_index)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef ptr;
+ LLVMValueRef res;
+ struct lp_type viewport_type =
+ lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
+
+ ptr = lp_jit_context_viewports(gallivm, context_ptr);
+ ptr = LLVMBuildPointerCast(builder, ptr,
+ LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
+
+ res = lp_build_pointer_get(builder, ptr, viewport_index);
+
+ return res;
+}
+
+
+static LLVMValueRef
+lp_build_depth_clamp(struct gallivm_state *gallivm,
+ LLVMBuilderRef builder,
+ struct lp_type type,
+ LLVMValueRef context_ptr,
+ LLVMValueRef thread_data_ptr,
+ LLVMValueRef z)
+{
+ LLVMValueRef viewport, min_depth, max_depth;
+ LLVMValueRef viewport_index;
+ struct lp_build_context f32_bld;
+
+ assert(type.floating);
+ lp_build_context_init(&f32_bld, gallivm, type);
+
+ /*
+ * Assumes clamping of the viewport index will occur in setup/gs. Value
+ * is passed through the rasterization stage via lp_rast_shader_inputs.
+ *
+ * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
+ * semantics.
+ */
+ viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
+ thread_data_ptr);
+
+ /*
+ * Load the min and max depth from the lp_jit_context.viewports
+ * array of lp_jit_viewport structures.
+ */
+ viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
+
+ /* viewports[viewport_index].min_depth */
+ min_depth = LLVMBuildExtractElement(builder, viewport,
+ lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
+ min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
+
+ /* viewports[viewport_index].max_depth */
+ max_depth = LLVMBuildExtractElement(builder, viewport,
+ lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
+ max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
+
+ /*
+ * Clamp to the min and max depth values for the given viewport.
+ */
+ return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
+}
+
+
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
*/
LLVMValueRef context_ptr,
LLVMValueRef num_loop,
struct lp_build_interp_soa_context *interp,
- struct lp_build_sampler_soa *sampler,
+ const struct lp_build_sampler_soa *sampler,
+ const struct lp_build_image_soa *image,
LLVMValueRef mask_store,
LLVMValueRef (*out_color)[4],
LLVMValueRef depth_ptr,
{
const struct util_format_description *zs_format_desc = NULL;
const struct tgsi_token *tokens = shader->base.tokens;
- LLVMTypeRef vec_type;
+ struct lp_type int_type = lp_int_type(type);
+ LLVMTypeRef vec_type, int_vec_type;
LLVMValueRef mask_ptr, mask_val;
- LLVMValueRef consts_ptr;
+ LLVMValueRef consts_ptr, num_consts_ptr;
+ LLVMValueRef ssbo_ptr, num_ssbo_ptr;
LLVMValueRef z;
- LLVMValueRef zs_value = NULL;
+ LLVMValueRef z_value, s_value;
+ LLVMValueRef z_fb, s_fb;
LLVMValueRef stencil_refs[2];
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
struct lp_build_for_loop_state loop_state;
struct lp_build_mask_context mask;
+ /*
+ * TODO: figure out if simple_shader optimization is really worthwile to
+ * keep. Disabled because it may hide some real bugs in the (depth/stencil)
+ * code since tests tend to take another codepath than real shaders.
+ */
boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
shader->info.base.num_inputs < 3 &&
- shader->info.base.num_instructions < 8);
+ shader->info.base.num_instructions < 8) && 0;
const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
util_blend_state_is_dual(&key->blend, 0);
unsigned attrib;
memset(&system_values, 0, sizeof(system_values));
if (key->depth.enabled ||
- key->stencil[0].enabled ||
- key->stencil[1].enabled) {
+ key->stencil[0].enabled) {
zs_format_desc = util_format_description(key->zsbuf_format);
assert(zs_format_desc);
- if (!shader->info.base.writes_z) {
- if (key->alpha.enabled || shader->info.base.uses_kill)
+ if (shader->info.base.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL])
+ depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
+ else if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
+ if (shader->info.base.writes_memory)
+ depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
+ else if (key->alpha.enabled ||
+ key->blend.alpha_to_coverage ||
+ shader->info.base.uses_kill ||
+ shader->info.base.writes_samplemask) {
/* With alpha test and kill, can do the depth test early
* and hopefully eliminate some quads. But need to do a
* special deferred depth write once the final mask value
- * is known.
+ * is known. This only works though if there's either no
+ * stencil test or the stencil value isn't written.
*/
- depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
+ if (key->stencil[0].enabled && (key->stencil[0].writemask ||
+ (key->stencil[1].enabled &&
+ key->stencil[1].writemask)))
+ depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
+ else
+ depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
+ }
else
depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
}
}
if (!(key->depth.enabled && key->depth.writemask) &&
- !(key->stencil[0].enabled && key->stencil[0].writemask))
+ !(key->stencil[0].enabled && (key->stencil[0].writemask ||
+ (key->stencil[1].enabled &&
+ key->stencil[1].writemask))))
depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
}
else {
depth_mode = 0;
}
+ vec_type = lp_build_vec_type(gallivm, type);
+ int_vec_type = lp_build_vec_type(gallivm, int_type);
stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
-
- vec_type = lp_build_vec_type(gallivm, type);
+ /* convert scalar stencil refs into vectors */
+ stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
+ stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
+ num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
+
+ ssbo_ptr = lp_jit_context_ssbos(gallivm, context_ptr);
+ num_ssbo_ptr = lp_jit_context_num_ssbos(gallivm, context_ptr);
lp_build_for_loop_begin(&loop_state, gallivm,
lp_build_const_int32(gallivm, 0),
z = interp->pos[2];
if (depth_mode & EARLY_DEPTH_TEST) {
- LLVMValueRef zs_dst_val;
- zs_dst_val = lp_build_depth_stencil_load_swizzled(gallivm, type,
- zs_format_desc,
- depth_ptr, depth_stride,
- loop_state.counter);
+ /*
+ * Clamp according to ARB_depth_clamp semantics.
+ */
+ if (key->depth_clamp) {
+ z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
+ thread_data_ptr, z);
+ }
+ lp_build_depth_stencil_load_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ depth_ptr, depth_stride,
+ &z_fb, &s_fb, loop_state.counter);
lp_build_depth_stencil_test(gallivm,
&key->depth,
key->stencil,
zs_format_desc,
&mask,
stencil_refs,
- z,
- zs_dst_val,
+ z, z_fb, s_fb,
facing,
- &zs_value,
+ &z_value, &s_value,
!simple_shader);
if (depth_mode & EARLY_DEPTH_WRITE) {
- lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
- NULL, loop_state.counter,
- depth_ptr, depth_stride, zs_value);
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ NULL, NULL, NULL, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
}
+ /*
+ * Note mask check if stencil is enabled must be after ds write not after
+ * stencil test otherwise new stencil values may not get written if all
+ * fragments got killed by depth/stencil test.
+ */
+ if (!simple_shader && key->stencil[0].enabled)
+ lp_build_mask_check(&mask);
}
lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
+ struct lp_build_tgsi_params params;
+ memset(¶ms, 0, sizeof(params));
+
+ params.type = type;
+ params.mask = &mask;
+ params.consts_ptr = consts_ptr;
+ params.const_sizes_ptr = num_consts_ptr;
+ params.system_values = &system_values;
+ params.inputs = interp->inputs;
+ params.context_ptr = context_ptr;
+ params.thread_data_ptr = thread_data_ptr;
+ params.sampler = sampler;
+ params.info = &shader->info.base;
+ params.ssbo_ptr = ssbo_ptr;
+ params.ssbo_sizes_ptr = num_ssbo_ptr;
+ params.image = image;
+
/* Build the actual shader */
- lp_build_tgsi_soa(gallivm, tokens, type, &mask,
- consts_ptr, &system_values,
- interp->inputs,
- outputs, sampler, &shader->info.base, NULL);
+ lp_build_tgsi_soa(gallivm, tokens, ¶ms,
+ outputs);
/* Alpha test */
if (key->alpha.enabled) {
}
}
+ /* Emulate Alpha to Coverage with Alpha test */
+ if (key->blend.alpha_to_coverage) {
+ int color0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_COLOR,
+ 0);
+
+ if (color0 != -1 && outputs[color0][3]) {
+ LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
+
+ lp_build_alpha_to_coverage(gallivm, type,
+ &mask, alpha,
+ (depth_mode & LATE_DEPTH_TEST) != 0);
+ }
+ }
+
+ if (shader->info.base.writes_samplemask) {
+ int smaski = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_SAMPLEMASK,
+ 0);
+ LLVMValueRef smask;
+ struct lp_build_context smask_bld;
+ lp_build_context_init(&smask_bld, gallivm, int_type);
+
+ assert(smaski >= 0);
+ smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
+ /*
+ * Pixel is alive according to the first sample in the mask.
+ */
+ smask = LLVMBuildBitCast(builder, smask, smask_bld.vec_type, "");
+ smask = lp_build_and(&smask_bld, smask, smask_bld.one);
+ smask = lp_build_cmp(&smask_bld, PIPE_FUNC_NOTEQUAL, smask, smask_bld.zero);
+ lp_build_mask_update(&mask, smask);
+ }
+
/* Late Z test */
if (depth_mode & LATE_DEPTH_TEST) {
- LLVMValueRef zs_dst_val;
int pos0 = find_output_by_semantic(&shader->info.base,
TGSI_SEMANTIC_POSITION,
0);
-
+ int s_out = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_STENCIL,
+ 0);
if (pos0 != -1 && outputs[pos0][2]) {
z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
}
+ /*
+ * Clamp according to ARB_depth_clamp semantics.
+ */
+ if (key->depth_clamp) {
+ z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
+ thread_data_ptr, z);
+ }
+
+ if (s_out != -1 && outputs[s_out][1]) {
+ /* there's only one value, and spec says to discard additional bits */
+ LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
+ stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
+ stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
+ stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
+ stencil_refs[1] = stencil_refs[0];
+ }
- zs_dst_val = lp_build_depth_stencil_load_swizzled(gallivm, type,
- zs_format_desc,
- depth_ptr, depth_stride,
- loop_state.counter);
+ lp_build_depth_stencil_load_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ depth_ptr, depth_stride,
+ &z_fb, &s_fb, loop_state.counter);
lp_build_depth_stencil_test(gallivm,
&key->depth,
zs_format_desc,
&mask,
stencil_refs,
- z,
- zs_dst_val,
+ z, z_fb, s_fb,
facing,
- &zs_value,
+ &z_value, &s_value,
!simple_shader);
/* Late Z write */
if (depth_mode & LATE_DEPTH_WRITE) {
- lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
- NULL, loop_state.counter,
- depth_ptr, depth_stride, zs_value);
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ NULL, NULL, NULL, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
}
}
else if ((depth_mode & EARLY_DEPTH_TEST) &&
* depth value, update from zs_value with the new mask value and
* write that out.
*/
- lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
- &mask, loop_state.counter,
- depth_ptr, depth_stride, zs_value);
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ &mask, z_fb, s_fb, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
}
*
* @param type fragment shader type (4x or 8x float)
* @param num_fs number of fs_src
+ * @param is_1d whether we're outputting to a 1d resource
* @param dst_channels number of output channels
* @param fs_src output from fragment shader
* @param dst pointer to store result
bool twiddle;
bool split;
- unsigned pixels = num_fs == 4 ? 1 : 2;
+ unsigned pixels = type.length / 4;
unsigned reorder_group;
unsigned src_channels;
unsigned src_count;
src_channels = dst_channels < 3 ? dst_channels : 4;
src_count = num_fs * src_channels;
- assert(pixels == 2 || num_fs == 4);
- assert(num_fs * src_channels <= Elements(src));
+ assert(pixels == 2 || pixels == 1);
+ assert(num_fs * src_channels <= ARRAY_SIZE(src));
/*
* Transpose from SoA -> AoS
}
} else if (twiddle) {
/* Twiddle pixels across elements of array */
+ /*
+ * XXX: we should avoid this in some cases, but would need to tell
+ * lp_build_conv to reorder (or deal with it ourselves).
+ */
lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
} else {
/* Do nothing */
}
+/*
+ * Untwiddle and transpose, much like the above.
+ * However, this is after conversion, so we get packed vectors.
+ * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
+ * the vectors will look like:
+ * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
+ * be swizzled here). Extending to 16bit should be trivial.
+ * Should also be extended to handle twice wide vectors with AVX2...
+ */
+static void
+fs_twiddle_transpose(struct gallivm_state *gallivm,
+ struct lp_type type,
+ LLVMValueRef *src,
+ unsigned src_count,
+ LLVMValueRef *dst)
+{
+ unsigned i, j;
+ struct lp_type type64, type16, type32;
+ LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef tmp[4], shuf[8];
+ for (j = 0; j < 2; j++) {
+ shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
+ shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
+ shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
+ shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
+ }
+
+ assert(src_count == 4 || src_count == 2 || src_count == 1);
+ assert(type.width == 8);
+ assert(type.length == 16);
+
+ type8_t = lp_build_vec_type(gallivm, type);
+
+ type64 = type;
+ type64.length /= 8;
+ type64.width *= 8;
+ type64_t = lp_build_vec_type(gallivm, type64);
+
+ type16 = type;
+ type16.length /= 2;
+ type16.width *= 2;
+ type16_t = lp_build_vec_type(gallivm, type16);
+
+ type32 = type;
+ type32.length /= 4;
+ type32.width *= 4;
+ type32_t = lp_build_vec_type(gallivm, type32);
+
+ lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
+
+ if (src_count == 1) {
+ /* transpose was no-op, just untwiddle */
+ LLVMValueRef shuf_vec;
+ shuf_vec = LLVMConstVector(shuf, 8);
+ tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
+ tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
+ dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
+ } else if (src_count == 2) {
+ LLVMValueRef shuf_vec;
+ shuf_vec = LLVMConstVector(shuf, 4);
+
+ for (i = 0; i < 2; i++) {
+ tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
+ tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
+ dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
+ }
+ } else {
+ for (j = 0; j < 2; j++) {
+ LLVMValueRef lo, hi, lo2, hi2;
+ /*
+ * Note that if we only really have 3 valid channels (rgb)
+ * and we don't need alpha we could substitute a undef here
+ * for the respective channel (causing llvm to drop conversion
+ * for alpha).
+ */
+ /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
+ lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
+ hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
+ lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
+ hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
+ dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
+ dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
+ }
+ }
+}
+
+
/**
* Load an unswizzled block of pixels from memory
*/
gep[1] = LLVMBuildAdd(builder, bx, by, "");
dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
- dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
+ dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
+ LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
- lp_set_load_alignment(dst[i], dst_alignment);
+ LLVMSetAlignment(dst[i], dst_alignment);
}
}
gep[1] = LLVMBuildAdd(builder, bx, by, "");
src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
- src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
+ src_ptr = LLVMBuildBitCast(builder, src_ptr,
+ LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
- lp_set_store_alignment(src_ptr, src_alignment);
+ LLVMSetAlignment(src_ptr, src_alignment);
}
}
*
* A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
*/
-static INLINE boolean
+static inline boolean
is_arithmetic_format(const struct util_format_description *format_desc)
{
boolean arith = false;
}
+/**
+ * Checks if this format requires special handling due to required expansion
+ * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
+ * SoA conversion.
+ */
+static inline boolean
+format_expands_to_float_soa(const struct util_format_description *format_desc)
+{
+ if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ return true;
+ }
+ return false;
+}
+
+
/**
* Retrieves the type representing the memory layout for a format
*
* e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
*/
-static INLINE void
+static inline void
lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
struct lp_type* type)
{
unsigned i;
unsigned chan;
- if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
- /* just make this a 32bit uint */
+ if (format_expands_to_float_soa(format_desc)) {
+ /* just make this a uint with width of block */
type->floating = false;
type->fixed = false;
type->sign = false;
type->norm = false;
- type->width = 32;
+ type->width = format_desc->block.bits;
type->length = 1;
return;
}
*
* e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
*/
-static INLINE void
+static inline void
lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
struct lp_type* type)
{
unsigned i;
unsigned chan;
- if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(format_desc)) {
/* always use ordinary floats for blending */
type->floating = true;
type->fixed = false;
/**
- * Scale a normalized value from src_bits to dst_bits
+ * Scale a normalized value from src_bits to dst_bits.
+ *
+ * The exact calculation is
+ *
+ * dst = iround(src * dst_mask / src_mask)
+ *
+ * or with integer rounding
+ *
+ * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
+ *
+ * where
+ *
+ * src_mask = (1 << src_bits) - 1
+ * dst_mask = (1 << dst_bits) - 1
+ *
+ * but we try to avoid division and multiplication through shifts.
*/
-static INLINE LLVMValueRef
+static inline LLVMValueRef
scale_bits(struct gallivm_state *gallivm,
int src_bits,
int dst_bits,
LLVMValueRef result = src;
if (dst_bits < src_bits) {
- /* Scale down by LShr */
- result = LLVMBuildLShr(builder,
- src,
- lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
- "");
+ int delta_bits = src_bits - dst_bits;
+
+ if (delta_bits <= dst_bits) {
+ /*
+ * Approximate the rescaling with a single shift.
+ *
+ * This gives the wrong rounding.
+ */
+
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+
+ } else {
+ /*
+ * Try more accurate rescaling.
+ */
+
+ /*
+ * Drop the least significant bits to make space for the multiplication.
+ *
+ * XXX: A better approach would be to use a wider integer type as intermediate. But
+ * this is enough to convert alpha from 16bits -> 2 when rendering to
+ * PIPE_FORMAT_R10G10B10A2_UNORM.
+ */
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, dst_bits),
+ "");
+
+
+ result = LLVMBuildMul(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
+ "");
+
+ /*
+ * Add a rounding term before the division.
+ *
+ * TODO: Handle signed integers too.
+ */
+ if (!src_type.sign) {
+ result = LLVMBuildAdd(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
+ "");
+ }
+
+ /*
+ * Approximate the division by src_mask with a src_bits shift.
+ *
+ * Given the src has already been shifted by dst_bits, all we need
+ * to do is to shift by the difference.
+ */
+
+ result = LLVMBuildLShr(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+ }
+
} else if (dst_bits > src_bits) {
/* Scale up bits */
int db = dst_bits - src_bits;
lp_build_const_int_vec(gallivm, src_type, db),
"");
- if (db < src_bits) {
+ if (db <= src_bits) {
/* Enough bits in src to fill the remainder */
LLVMValueRef lower = LLVMBuildLShr(builder,
src,
return result;
}
+/**
+ * If RT is a smallfloat (needing denorms) format
+ */
+static inline int
+have_smallfloat_format(struct lp_type dst_type,
+ enum pipe_format format)
+{
+ return ((dst_type.floating && dst_type.width != 32) ||
+ /* due to format handling hacks this format doesn't have floating set
+ * here (and actually has width set to 32 too) so special case this. */
+ (format == PIPE_FORMAT_R11G11B10_FLOAT));
+}
+
/**
* Convert from memory format to blending format
*/
static void
convert_to_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
const struct util_format_description *src_fmt,
struct lp_type src_type,
struct lp_type dst_type,
LLVMBuilderRef builder = gallivm->builder;
struct lp_type blend_type;
struct lp_type mem_type;
- unsigned i, j, k;
- unsigned pixels = 16 / num_srcs;
+ unsigned i, j;
+ unsigned pixels = block_size / num_srcs;
bool is_arith;
/*
- * full custom path for packed floats - none of the later functions would do
- * anything useful, and given the lp_type representation they can't be fixed.
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
*/
- if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(src_fmt)) {
LLVMValueRef tmpsrc[4];
/*
* This is pretty suboptimal for this case blending in SoA would be much
* better, since conversion gets us SoA values so need to convert back.
*/
- assert(src_type.width == 32);
+ assert(src_type.width == 32 || src_type.width == 16);
assert(dst_type.floating);
assert(dst_type.width == 32);
assert(dst_type.length % 4 == 0);
+ assert(num_srcs % 4 == 0);
+
+ if (src_type.width == 16) {
+ /* expand 4x16bit values to 4x32bit */
+ struct lp_type type32x4 = src_type;
+ LLVMTypeRef ltype32x4;
+ unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
+ type32x4.width = 32;
+ ltype32x4 = lp_build_vec_type(gallivm, type32x4);
+ for (i = 0; i < num_fetch; i++) {
+ src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
+ }
+ src_type.width = 32;
+ }
for (i = 0; i < 4; i++) {
tmpsrc[i] = src[i];
}
for (i = 0; i < num_srcs / 4; i++) {
LLVMValueRef tmpsoa[4];
LLVMValueRef tmps = tmpsrc[i];
- if (num_srcs == 8) {
+ if (dst_type.length == 8) {
LLVMValueRef shuffles[8];
unsigned j;
/* fetch was 4 values but need 8-wide output values */
tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
LLVMConstVector(shuffles, 8), "");
}
- lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
+ }
+ else {
+ lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
+ }
lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
}
return;
for (i = 0; i < num_srcs; ++i) {
LLVMValueRef chans[4];
LLVMValueRef res = NULL;
- unsigned sa = 0;
dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#if PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
- for (k = 0; k < src_fmt->channel[j].size; ++k) {
- mask |= 1 << k;
- }
+ mask = (1 << src_fmt->channel[j].size) - 1;
/* Extract bits from source */
chans[j] = LLVMBuildLShr(builder,
/* Insert bits into correct position */
chans[j] = LLVMBuildShl(builder,
chans[j],
- lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
+ lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
"");
- sa += src_fmt->channel[j].size;
-
if (j == 0) {
res = chans[j];
} else {
*/
static void
convert_from_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
const struct util_format_description *src_fmt,
struct lp_type src_type,
struct lp_type dst_type,
struct lp_type mem_type;
struct lp_type blend_type;
LLVMBuilderRef builder = gallivm->builder;
- unsigned pixels = 16 / num_srcs;
+ unsigned pixels = block_size / num_srcs;
bool is_arith;
/*
- * full custom path for packed floats - none of the later functions would do
- * anything useful, and given the lp_type representation they can't be fixed.
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
*/
- if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(src_fmt)) {
/*
* This is pretty suboptimal for this case blending in SoA would be much
* better - we need to transpose the AoS values back to SoA values for
assert(src_type.floating);
assert(src_type.width == 32);
assert(src_type.length % 4 == 0);
- assert(dst_type.width == 32);
+ assert(dst_type.width == 32 || dst_type.width == 16);
+
for (i = 0; i < num_srcs / 4; i++) {
LLVMValueRef tmpsoa[4], tmpdst;
lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
- tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
- if (num_srcs == 8) {
+ /* really really need SoA here */
+
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
+ }
+ else {
+ tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
+ src_type, tmpsoa);
+ }
+
+ if (src_type.length == 8) {
LLVMValueRef tmpaos, shuffles[8];
unsigned j;
/*
src[i] = tmpdst;
}
}
+ if (dst_type.width == 16) {
+ struct lp_type type16x8 = dst_type;
+ struct lp_type type32x4 = dst_type;
+ LLVMTypeRef ltype16x4, ltypei64, ltypei128;
+ unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
+ type16x8.length = 8;
+ type32x4.width = 32;
+ ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
+ ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
+ ltype16x4 = lp_build_vec_type(gallivm, dst_type);
+ /* We could do vector truncation but it doesn't generate very good code */
+ for (i = 0; i < num_fetch; i++) {
+ src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
+ src[i], lp_build_zero(gallivm, type32x4));
+ src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
+ src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
+ src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
+ }
+ }
return;
}
for (i = 0; i < num_srcs; ++i) {
LLVMValueRef chans[4];
LLVMValueRef res = NULL;
- unsigned sa = 0;
dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#if PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
assert(blend_type.width > src_fmt->channel[j].size);
/* Extract bits */
chans[j] = LLVMBuildLShr(builder,
dst[i],
- lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
+ lp_build_const_int_vec(gallivm, src_type,
+ from_lsb * blend_type.width),
"");
chans[j] = LLVMBuildAnd(builder,
row_type.length = alpha_type.length;
/* Twiddle the alpha to match pixels */
- lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
+ lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
- for (i = 0; i < 4; ++i) {
+ /*
+ * TODO this should use single lp_build_conv call for
+ * src_count == 1 && dst_channels == 1 case (dropping the concat below)
+ */
+ for (i = 0; i < block_height; ++i) {
lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
}
row_type.length = length;
/* If only one channel we can only need the single alpha value per pixel */
- if (src_count == 1) {
- assert(dst_channels == 1);
+ if (src_count == 1 && dst_channels == 1) {
- lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
+ lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
} else {
/* If there are more srcs than rows then we need to split alpha up */
if (src_count > block_height) {
unsigned pixels = block_size / src_count;
unsigned idx = i - 1;
- src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
+ src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
}
}
/* If there is a src for each pixel broadcast the alpha across whole row */
if (src_count == block_size) {
for (i = 0; i < src_count; ++i) {
- src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
+ src_alpha[i] = lp_build_broadcast(gallivm,
+ lp_build_vec_type(gallivm, row_type), src_alpha[i]);
}
} else {
unsigned pixels = block_size / src_count;
boolean do_branch)
{
const unsigned alpha_channel = 3;
- const unsigned block_width = 4;
- const unsigned block_height = 4;
+ const unsigned block_width = LP_RASTER_BLOCK_SIZE;
+ const unsigned block_height = LP_RASTER_BLOCK_SIZE;
const unsigned block_size = block_width * block_height;
const unsigned lp_integer_vector_width = 128;
LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
LLVMValueRef src_alpha[4 * 4];
- LLVMValueRef src1_alpha[4 * 4];
+ LLVMValueRef src1_alpha[4 * 4] = { NULL };
LLVMValueRef src_mask[4 * 4];
LLVMValueRef src[4 * 4];
LLVMValueRef src1[4 * 4];
LLVMValueRef blend_alpha;
LLVMValueRef i32_zero;
LLVMValueRef check_mask;
+ LLVMValueRef undef_src_val;
struct lp_build_mask_context mask_ctx;
struct lp_type mask_type;
struct lp_type blend_type;
struct lp_type row_type;
struct lp_type dst_type;
+ struct lp_type ls_type;
unsigned char swizzle[TGSI_NUM_CHANNELS];
unsigned vector_width;
const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
util_blend_state_is_dual(&variant->key.blend, 0);
+ const boolean is_1d = variant->key.resource_1d;
+ boolean twiddle_after_convert = FALSE;
+ unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
+ LLVMValueRef fpstate = 0;
+
+ /* Get type from output format */
+ lp_blend_type_from_format_desc(out_format_desc, &row_type);
+ lp_mem_type_from_format_desc(out_format_desc, &dst_type);
+
+ /*
+ * Technically this code should go into lp_build_smallfloat_to_float
+ * and lp_build_float_to_smallfloat but due to the
+ * http://llvm.org/bugs/show_bug.cgi?id=6393
+ * llvm reorders the mxcsr intrinsics in a way that breaks the code.
+ * So the ordering is important here and there shouldn't be any
+ * llvm ir instrunctions in this function before
+ * this, otherwise half-float format conversions won't work
+ * (again due to llvm bug #6393).
+ */
+ if (have_smallfloat_format(dst_type, out_format)) {
+ /* We need to make sure that denorms are ok for half float
+ conversions */
+ fpstate = lp_build_fpstate_get(gallivm);
+ lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
+ }
+
mask_type = lp_int32_vec4_type();
mask_type.length = fs_type.length;
- /* Compute the alignment of the destination pointer in bytes */
-#if 0
- dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
-#else
- /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
- dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
-#endif
- /* Force power-of-two alignment by extracting only the least-significant-bit */
- dst_alignment = 1 << (ffs(dst_alignment) - 1);
- /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
- dst_alignment = MIN2(dst_alignment, 16);
+ for (i = num_fs; i < num_fullblock_fs; i++) {
+ fs_mask[i] = lp_build_zero(gallivm, mask_type);
+ }
/* Do not bother executing code when mask is empty.. */
if (do_branch) {
check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
}
partial_mask |= !variant->opaque;
i32_zero = lp_build_const_int32(gallivm, 0);
- /* Get type from output format */
- lp_blend_type_from_format_desc(out_format_desc, &row_type);
- lp_mem_type_from_format_desc(out_format_desc, &dst_type);
+ undef_src_val = lp_build_undef(gallivm, fs_type);
row_type.length = fs_type.length;
vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
}
}
- if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
- /* the code above can't work for layout_other */
+ if (format_expands_to_float_soa(out_format_desc)) {
+ /*
+ * the code above can't work for layout_other
+ * for srgb it would sort of work but we short-circuit swizzles, etc.
+ * as that is done as part of unpack / pack.
+ */
dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
has_alpha = true;
swizzle[0] = 0;
}
/* If 3 channels then pad to include alpha for 4 element transpose */
- if (dst_channels == 3 && !has_alpha) {
+ if (dst_channels == 3) {
+ assert (!has_alpha);
for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
if (swizzle[i] > TGSI_NUM_CHANNELS)
swizzle[i] = 3;
}
if (out_format_desc->nr_channels == 4) {
dst_channels = 4;
+ /*
+ * We use alpha from the color conversion, not separate one.
+ * We had to include it for transpose, hence it will get converted
+ * too (albeit when doing transpose after conversion, that would
+ * no longer be the case necessarily).
+ * (It works only with 4 channel dsts, e.g. rgbx formats, because
+ * otherwise we really have padding, not alpha, included.)
+ */
+ has_alpha = true;
}
}
/*
* Load shader output
*/
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
/* Always load alpha for use in blending */
- LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
/* Load each channel */
for (j = 0; j < dst_channels; ++j) {
assert(swizzle[j] < 4);
- fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
+ if (i < num_fs) {
+ fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
+ }
+ else {
+ fs_src[i][j] = undef_src_val;
+ }
}
/* If 3 channels then pad to include alpha for 4 element transpose */
/*
* XXX If we include that here maybe could actually use it instead of
* separate alpha for blending?
+ * (Difficult though we actually convert pad channels, not alpha.)
*/
if (dst_channels == 3 && !has_alpha) {
fs_src[i][3] = alpha;
/* We split the row_mask and row_alpha as we want 128bit interleave */
if (fs_type.length == 8) {
- src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
- src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
+ src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],
+ 0, src_channels);
+ src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],
+ src_channels, src_channels);
src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
- src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
+ src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
+ src_channels, src_channels);
} else {
src_mask[i] = fs_mask[i];
src_alpha[i] = alpha;
}
if (dual_source_blend) {
/* same as above except different src/dst, skip masks and comments... */
- for (i = 0; i < num_fs; ++i) {
- LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
+ for (i = 0; i < num_fullblock_fs; ++i) {
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
for (j = 0; j < dst_channels; ++j) {
assert(swizzle[j] < 4);
- fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
+ if (i < num_fs) {
+ fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
+ }
+ else {
+ fs_src1[i][j] = undef_src_val;
+ }
}
if (dst_channels == 3 && !has_alpha) {
fs_src1[i][3] = alpha;
}
if (fs_type.length == 8) {
src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
- src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
+ src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
+ src_channels, src_channels);
} else {
src1_alpha[i] = alpha;
}
*/
fs_type.floating = 0;
fs_type.sign = dst_type.sign;
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
for (j = 0; j < dst_channels; ++j) {
fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
lp_build_vec_type(gallivm, fs_type), "");
}
}
+ /*
+ * We actually should generally do conversion first (for non-1d cases)
+ * when the blend format is 8 or 16 bits. The reason is obvious,
+ * there's 2 or 4 times less vectors to deal with for the interleave...
+ * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
+ * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
+ * unpack only with 128bit vectors).
+ * Note: for 16bit sizes really need matching pack conversion code
+ */
+ if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
+ twiddle_after_convert = TRUE;
+ }
+
/*
* Pixel twiddle from fragment shader order to memory order
*/
- src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
- if (dual_source_blend) {
- generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
+ if (!twiddle_after_convert) {
+ src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
+ dst_channels, fs_src, src, pad_inline);
+ if (dual_source_blend) {
+ generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
+ fs_src1, src1, pad_inline);
+ }
+ } else {
+ src_count = num_fullblock_fs * dst_channels;
+ /*
+ * We reorder things a bit here, so the cases for 4-wide and 8-wide
+ * (AVX) turn out the same later when untwiddling/transpose (albeit
+ * for true AVX2 path untwiddle needs to be different).
+ * For now just order by colors first (so we can use unpack later).
+ */
+ for (j = 0; j < num_fullblock_fs; j++) {
+ for (i = 0; i < dst_channels; i++) {
+ src[i*num_fullblock_fs + j] = fs_src[j][i];
+ if (dual_source_blend) {
+ src1[i*num_fullblock_fs + j] = fs_src1[j][i];
+ }
+ }
+ }
}
src_channels = dst_channels < 3 ? dst_channels : 4;
- if (src_count != num_fs * src_channels) {
- unsigned ds = src_count / (num_fs * src_channels);
+ if (src_count != num_fullblock_fs * src_channels) {
+ unsigned ds = src_count / (num_fullblock_fs * src_channels);
row_type.length /= ds;
fs_type.length = row_type.length;
}
unsigned bits = row_type.width * row_type.length;
unsigned combined;
+ assert(src_count >= (vector_width / bits));
+
dst_count = src_count / (vector_width / bits);
+
combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
if (dual_source_blend) {
lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
assert(bits == 128 || bits == 256);
}
+ if (twiddle_after_convert) {
+ fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
+ if (dual_source_blend) {
+ fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
+ }
+ }
/*
* Blend Colour conversion
*/
blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
- blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
- blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
+ blend_color = LLVMBuildPointerCast(builder, blend_color,
+ LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
+ blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
+ &i32_zero, 1, ""), "");
/* Convert */
lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
+ if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ /*
+ * since blending is done with floats, there was no conversion.
+ * However, the rules according to fixed point renderbuffers still
+ * apply, that is we must clamp inputs to 0.0/1.0.
+ * (This would apply to separate alpha conversion too but we currently
+ * force has_alpha to be true.)
+ * TODO: should skip this with "fake" blend, since post-blend conversion
+ * will clamp anyway.
+ * TODO: could also skip this if fragment color clamping is enabled. We
+ * don't support it natively so it gets baked into the shader however, so
+ * can't really tell here.
+ */
+ struct lp_build_context f32_bld;
+ assert(row_type.floating);
+ lp_build_context_init(&f32_bld, gallivm, row_type);
+ for (i = 0; i < src_count; i++) {
+ src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
+ }
+ if (dual_source_blend) {
+ for (i = 0; i < src_count; i++) {
+ src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
+ }
+ }
+ /* probably can't be different than row_type but better safe than sorry... */
+ lp_build_context_init(&f32_bld, gallivm, blend_type);
+ blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
+ }
+
/* Extract alpha */
blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
/*
* Mask conversion
*/
- lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
+ lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
if (src_count < block_height) {
lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
unsigned pixels = block_size / src_count;
unsigned idx = i - 1;
- src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
+ src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
}
}
mask_type.length = pixels;
mask_type.width = row_type.width * dst_channels;
- src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
+ /*
+ * If mask_type width is smaller than 32bit, this doesn't quite
+ * generate the most efficient code (could use some pack).
+ */
+ src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
+ lp_build_int_vec_type(gallivm, mask_type), "");
mask_type.length *= dst_channels;
mask_type.width /= dst_channels;
}
- src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
+ src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
+ lp_build_int_vec_type(gallivm, mask_type), "");
src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
}
dst_count = src_count;
}
- dst_type.length *= 16 / dst_count;
+ dst_type.length *= block_size / dst_count;
- if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(out_format_desc)) {
/*
* we need multiple values at once for the conversion, so can as well
* load them vectorized here too instead of concatenating later.
dst_type.length = block_width;
}
- load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
- dst, dst_type, dst_count, dst_alignment);
+ /*
+ * Compute the alignment of the destination pointer in bytes
+ * We fetch 1-4 pixels, if the format has pot alignment then those fetches
+ * are always aligned by MIN2(16, fetch_width) except for buffers (not
+ * 1d tex but can't distinguish here) so need to stick with per-pixel
+ * alignment in this case.
+ */
+ if (is_1d) {
+ dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
+ }
+ else {
+ dst_alignment = dst_type.length * dst_type.width / 8;
+ }
+ /* Force power-of-two alignment by extracting only the least-significant-bit */
+ dst_alignment = 1 << (ffs(dst_alignment) - 1);
+ /*
+ * Resource base and stride pointers are aligned to 16 bytes, so that's
+ * the maximum alignment we can guarantee
+ */
+ dst_alignment = MIN2(16, dst_alignment);
+
+ ls_type = dst_type;
+
+ if (dst_count > src_count) {
+ if ((dst_type.width == 8 || dst_type.width == 16) &&
+ util_is_power_of_two_or_zero(dst_type.length) &&
+ dst_type.length * dst_type.width < 128) {
+ /*
+ * Never try to load values as 4xi8 which we will then
+ * concatenate to larger vectors. This gives llvm a real
+ * headache (the problem is the type legalizer (?) will
+ * try to load that as 4xi8 zext to 4xi32 to fill the vector,
+ * then the shuffles to concatenate are more or less impossible
+ * - llvm is easily capable of generating a sequence of 32
+ * pextrb/pinsrb instructions for that. Albeit it appears to
+ * be fixed in llvm 4.0. So, load and concatenate with 32bit
+ * width to avoid the trouble (16bit seems not as bad, llvm
+ * probably recognizes the load+shuffle as only one shuffle
+ * is necessary, but we can do just the same anyway).
+ */
+ ls_type.length = dst_type.length * dst_type.width / 32;
+ ls_type.width = 32;
+ }
+ }
+
+ if (is_1d) {
+ load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
+ dst, ls_type, dst_count / 4, dst_alignment);
+ for (i = dst_count / 4; i < dst_count; i++) {
+ dst[i] = lp_build_undef(gallivm, ls_type);
+ }
+
+ }
+ else {
+ load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
+ dst, ls_type, dst_count, dst_alignment);
+ }
/*
* on all 16 pixels in that single vector at once.
*/
if (dst_count > src_count) {
- lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
+ if (ls_type.length != dst_type.length && ls_type.length == 1) {
+ LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
+ LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
+ for (i = 0; i < dst_count; i++) {
+ dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
+ }
+ }
+
+ lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
+
+ if (ls_type.length != dst_type.length) {
+ struct lp_type tmp_type = dst_type;
+ tmp_type.length = dst_type.length * 4 / src_count;
+ for (i = 0; i < src_count; i++) {
+ dst[i] = LLVMBuildBitCast(builder, dst[i],
+ lp_build_vec_type(gallivm, tmp_type), "");
+ }
+ }
}
/*
* It seems some cleanup could be done here (like skipping conversion/blend
* when not needed).
*/
- convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
+ convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
+ row_type, dst, src_count);
+ /*
+ * FIXME: Really should get logic ops / masks out of generic blend / row
+ * format. Logic ops will definitely not work on the blend float format
+ * used for SRGB here and I think OpenGL expects this to work as expected
+ * (that is incoming values converted to srgb then logic op applied).
+ */
for (i = 0; i < src_count; ++i) {
dst[i] = lp_build_blend_aos(gallivm,
&variant->key.blend,
pad_inline ? 4 : dst_channels);
}
- convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
+ convert_from_blend_type(gallivm, block_size, out_format_desc,
+ row_type, dst_type, dst, src_count);
/* Split the blend rows back to memory rows */
if (dst_count > src_count) {
src_count *= 2;
}
-
/*
* Store blend result to memory
*/
- store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
- dst, dst_type, dst_count, dst_alignment);
+ if (is_1d) {
+ store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
+ dst, dst_type, dst_count / 4, dst_alignment);
+ }
+ else {
+ store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
+ dst, dst_type, dst_count, dst_alignment);
+ }
+
+ if (have_smallfloat_format(dst_type, out_format)) {
+ lp_build_fpstate_set(gallivm, fpstate);
+ }
if (do_branch) {
lp_build_mask_end(&mask_ctx);
unsigned partial_mask)
{
struct gallivm_state *gallivm = variant->gallivm;
- const struct lp_fragment_shader_variant_key *key = &variant->key;
+ struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
- char func_name[256];
+ char func_name[64];
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
struct lp_build_sampler_soa *sampler;
+ struct lp_build_image_soa *image;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[16 / 4];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
}
/* check if writes to cbuf[0] are to be copied to all cbufs */
- cbuf0_write_all = FALSE;
- for (i = 0;i < shader->info.base.num_properties; i++) {
- if (shader->info.base.properties[i].name ==
- TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
- cbuf0_write_all = TRUE;
- break;
- }
- }
+ cbuf0_write_all =
+ shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
- num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_vec_type = lp_build_vec_type(gallivm, blend_type);
- util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
- shader->no, variant->no, partial_mask ? "partial" : "whole");
+ snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
+ shader->no, variant->no, partial_mask ? "partial" : "whole");
arg_types[0] = variant->jit_context_ptr_type; /* context */
arg_types[1] = int32_type; /* x */
arg_types[12] = int32_type; /* depth_stride */
func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
- arg_types, Elements(arg_types), 0);
+ arg_types, ARRAY_SIZE(arg_types), 0);
function = LLVMAddFunction(gallivm->module, func_name, func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
- for(i = 0; i < Elements(arg_types); ++i)
+ for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
- LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
+ lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr_ptr");
lp_build_name(depth_ptr, "depth");
- lp_build_name(thread_data_ptr, "thread_data");
lp_build_name(mask_input, "mask_input");
+ lp_build_name(thread_data_ptr, "thread_data");
lp_build_name(stride_ptr, "stride_ptr");
lp_build_name(depth_stride, "depth_stride");
assert(builder);
LLVMPositionBuilderAtEnd(builder, block);
+ /*
+ * Must not count ps invocations if there's a null shader.
+ * (It would be ok to count with null shader if there's d/s tests,
+ * but only if there's d/s buffers too, which is different
+ * to implicit rasterization disable which must not depend
+ * on the d/s buffers.)
+ * Could use popcount on mask, but pixel accuracy is not required.
+ * Could disable if there's no stats query, but maybe not worth it.
+ */
+ if (shader->info.base.num_instructions > 1) {
+ LLVMValueRef invocs, val;
+ invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
+ val = LLVMBuildLoad(builder, invocs, "");
+ val = LLVMBuildAdd(builder, val,
+ LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
+ "invoc_count");
+ LLVMBuildStore(builder, val, invocs);
+ }
+
/* code generated texture sampling */
- sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
+ sampler = lp_llvm_sampler_soa_create(key->samplers);
+ image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key));
+
+ num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
+ /* for 1d resources only run "upper half" of stamp */
+ if (key->resource_1d)
+ num_fs /= 2;
{
LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
num_loop, "mask_store");
LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
+ boolean pixel_center_integer =
+ shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
/*
* The shader input interpolation info is not explicitely baked in the
gallivm,
shader->info.base.num_inputs,
inputs,
- shader->info.base.pixel_center_integer,
+ pixel_center_integer,
+ key->depth_clamp,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
num_loop,
&interp,
sampler,
+ image,
mask_store, /* output */
color_store,
depth_ptr,
}
sampler->destroy(sampler);
-
+ image->destroy(image);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
- LLVMValueRef color_ptr;
- LLVMValueRef stride;
- LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
+ if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
+ LLVMValueRef color_ptr;
+ LLVMValueRef stride;
+ LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
- boolean do_branch = ((key->depth.enabled
- || key->stencil[0].enabled
- || key->alpha.enabled)
- && !shader->info.base.uses_kill);
+ boolean do_branch = ((key->depth.enabled
+ || key->stencil[0].enabled
+ || key->alpha.enabled)
+ && !shader->info.base.uses_kill);
- color_ptr = LLVMBuildLoad(builder,
- LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
- "");
+ color_ptr = LLVMBuildLoad(builder,
+ LLVMBuildGEP(builder, color_ptr_ptr,
+ &index, 1, ""),
+ "");
- lp_build_name(color_ptr, "color_ptr%d", cbuf);
+ lp_build_name(color_ptr, "color_ptr%d", cbuf);
- stride = LLVMBuildLoad(builder,
- LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
- "");
+ stride = LLVMBuildLoad(builder,
+ LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
+ "");
- generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
- num_fs, fs_type, fs_mask, fs_out_color,
- context_ptr, color_ptr, stride, partial_mask, do_branch);
+ generate_unswizzled_blend(gallivm, cbuf, variant,
+ key->cbuf_format[cbuf],
+ num_fs, fs_type, fs_mask, fs_out_color,
+ context_ptr, color_ptr, stride,
+ partial_mask, do_branch);
+ }
}
LLVMBuildRetVoid(builder);
gallivm_verify_function(gallivm, function);
-
- variant->nr_instrs += lp_build_count_instructions(function);
}
static void
-dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
+dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
{
unsigned i;
for (i = 0; i < key->nr_cbufs; ++i) {
debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
}
- if (key->depth.enabled) {
+ if (key->depth.enabled || key->stencil[0].enabled) {
debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
- debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
+ }
+ if (key->depth.enabled) {
+ debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
debug_printf("depth.writemask = %u\n", key->depth.writemask);
}
for (i = 0; i < 2; ++i) {
if (key->stencil[i].enabled) {
- debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
- debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
- debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
- debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
+ debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
+ debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
+ debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
+ debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
}
}
if (key->alpha.enabled) {
- debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
+ debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
}
if (key->occlusion_count) {
}
if (key->blend.logicop_enable) {
- debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
+ debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
}
else if (key->blend.rt[0].blend_enable) {
- debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE));
- debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
- debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
- debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE));
- debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
- debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
+ debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));
+ debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
+ debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
+ debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));
+ debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
+ debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
}
debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
+ if (key->blend.alpha_to_coverage) {
+ debug_printf("blend.alpha_to_coverage is enabled\n");
+ }
for (i = 0; i < key->nr_samplers; ++i) {
- const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
+ const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
debug_printf("sampler[%u] = \n", i);
debug_printf(" .wrap = %s %s %s\n",
- util_dump_tex_wrap(sampler->wrap_s, TRUE),
- util_dump_tex_wrap(sampler->wrap_t, TRUE),
- util_dump_tex_wrap(sampler->wrap_r, TRUE));
+ util_str_tex_wrap(sampler->wrap_s, TRUE),
+ util_str_tex_wrap(sampler->wrap_t, TRUE),
+ util_str_tex_wrap(sampler->wrap_r, TRUE));
debug_printf(" .min_img_filter = %s\n",
- util_dump_tex_filter(sampler->min_img_filter, TRUE));
+ util_str_tex_filter(sampler->min_img_filter, TRUE));
debug_printf(" .min_mip_filter = %s\n",
- util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
+ util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
debug_printf(" .mag_img_filter = %s\n",
- util_dump_tex_filter(sampler->mag_img_filter, TRUE));
+ util_str_tex_filter(sampler->mag_img_filter, TRUE));
if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
- debug_printf(" .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
+ debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
}
for (i = 0; i < key->nr_sampler_views; ++i) {
- const struct lp_static_texture_state *texture = &key->state[i].texture_state;
+ const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
debug_printf("texture[%u] = \n", i);
debug_printf(" .format = %s\n",
util_format_name(texture->format));
debug_printf(" .target = %s\n",
- util_dump_tex_target(texture->target, TRUE));
+ util_str_tex_target(texture->target, TRUE));
debug_printf(" .level_zero_only = %u\n",
texture->level_zero_only);
debug_printf(" .pot = %u %u %u\n",
texture->pot_height,
texture->pot_depth);
}
+ struct lp_image_static_state *images = lp_fs_variant_key_images(key);
+ for (i = 0; i < key->nr_images; ++i) {
+ const struct lp_static_texture_state *image = &images[i].image_state;
+ debug_printf("image[%u] = \n", i);
+ debug_printf(" .format = %s\n",
+ util_format_name(image->format));
+ debug_printf(" .target = %s\n",
+ util_str_tex_target(image->target, TRUE));
+ debug_printf(" .level_zero_only = %u\n",
+ image->level_zero_only);
+ debug_printf(" .pot = %u %u %u\n",
+ image->pot_width,
+ image->pot_height,
+ image->pot_depth);
+ }
}
void
-lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
+lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
{
debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
variant->shader->no, variant->no);
const struct lp_fragment_shader_variant_key *key)
{
struct lp_fragment_shader_variant *variant;
- const struct util_format_description *cbuf0_format_desc;
+ const struct util_format_description *cbuf0_format_desc = NULL;
boolean fullcolormask;
+ char module_name[64];
- variant = CALLOC_STRUCT(lp_fragment_shader_variant);
- if(!variant)
+ variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
+ if (!variant)
return NULL;
- variant->gallivm = gallivm_create();
+ memset(variant, 0, sizeof(*variant));
+ snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
+ shader->no, shader->variants_created);
+
+ variant->gallivm = gallivm_create(module_name, lp->context);
if (!variant->gallivm) {
FREE(variant);
return NULL;
fullcolormask &&
!key->stencil[0].enabled &&
!key->alpha.enabled &&
+ !key->blend.alpha_to_coverage &&
!key->depth.enabled &&
- !shader->info.base.uses_kill
- ? TRUE : FALSE;
+ !shader->info.base.uses_kill &&
+ !shader->info.base.writes_samplemask
+ ? TRUE : FALSE;
if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
lp_debug_fs_variant(variant);
gallivm_compile_module(variant->gallivm);
+ variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
+
if (variant->function[RAST_EDGE_TEST]) {
variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
gallivm_jit_function(variant->gallivm,
variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
}
+ gallivm_free_ir(variant->gallivm);
+
return variant;
}
struct lp_fragment_shader *shader;
int nr_samplers;
int nr_sampler_views;
+ int nr_images;
int i;
shader = CALLOC_STRUCT(lp_fragment_shader);
nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
-
- shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
- state[MAX2(nr_samplers, nr_sampler_views)]);
+ nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
+ shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
for (i = 0; i < shader->info.base.num_inputs; i++) {
shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
switch (shader->info.base.input_interpolate[i]) {
case TGSI_INTERPOLATE_CONSTANT:
- shader->inputs[i].interp = LP_INTERP_CONSTANT;
- break;
+ shader->inputs[i].interp = LP_INTERP_CONSTANT;
+ break;
case TGSI_INTERPOLATE_LINEAR:
- shader->inputs[i].interp = LP_INTERP_LINEAR;
- break;
+ shader->inputs[i].interp = LP_INTERP_LINEAR;
+ break;
case TGSI_INTERPOLATE_PERSPECTIVE:
- shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
- break;
+ shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
+ break;
case TGSI_INTERPOLATE_COLOR:
- shader->inputs[i].interp = LP_INTERP_COLOR;
- break;
+ shader->inputs[i].interp = LP_INTERP_COLOR;
+ break;
default:
- assert(0);
- break;
+ assert(0);
+ break;
}
switch (shader->info.base.input_semantic_name[i]) {
case TGSI_SEMANTIC_FACE:
- shader->inputs[i].interp = LP_INTERP_FACING;
- break;
+ shader->inputs[i].interp = LP_INTERP_FACING;
+ break;
case TGSI_SEMANTIC_POSITION:
- /* Position was already emitted above
- */
- shader->inputs[i].interp = LP_INTERP_POSITION;
- shader->inputs[i].src_index = 0;
- continue;
+ /* Position was already emitted above
+ */
+ shader->inputs[i].interp = LP_INTERP_POSITION;
+ shader->inputs[i].src_index = 0;
+ continue;
}
+ /* XXX this is a completely pointless index map... */
shader->inputs[i].src_index = i+1;
}
* Remove shader variant from two lists: the shader's variant list
* and the context's variant list.
*/
-void
+static void
llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
struct lp_fragment_shader_variant *variant)
{
- unsigned i;
-
- if (gallivm_debug & GALLIVM_DEBUG_IR) {
- debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
- " #%u v total cached #%u\n",
- variant->shader->no,
- variant->no,
+ if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
+ debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
+ "v total cached %u inst %u total inst %u\n",
+ variant->shader->no, variant->no,
variant->shader->variants_created,
variant->shader->variants_cached,
- lp->nr_fs_variants);
- }
-
- /* free all the variant's JIT'd functions */
- for (i = 0; i < Elements(variant->function); i++) {
- if (variant->function[i]) {
- gallivm_free_function(variant->gallivm,
- variant->function[i],
- variant->jit_function[i]);
- }
+ lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);
}
gallivm_destroy(variant->gallivm);
static void
llvmpipe_set_constant_buffer(struct pipe_context *pipe,
- uint shader, uint index,
- struct pipe_constant_buffer *cb)
+ enum pipe_shader_type shader, uint index,
+ const struct pipe_constant_buffer *cb)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
struct pipe_resource *constants = cb ? cb->buffer : NULL;
assert(shader < PIPE_SHADER_TYPES);
- assert(index < Elements(llvmpipe->constants[shader]));
+ assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
/* note: reference counting */
util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
+ if (constants) {
+ if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
+ debug_printf("Illegal set constant without bind flag\n");
+ constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
+ }
+ }
+
if (shader == PIPE_SHADER_VERTEX ||
shader == PIPE_SHADER_GEOMETRY) {
/* Pass the constants to the 'draw' module */
draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
index, data, size);
}
-
- llvmpipe->dirty |= LP_NEW_CONSTANTS;
+ else if (shader == PIPE_SHADER_COMPUTE)
+ llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;
+ else
+ llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
if (cb && cb->user_buffer) {
pipe_resource_reference(&constants, NULL);
}
}
+static void
+llvmpipe_set_shader_buffers(struct pipe_context *pipe,
+ enum pipe_shader_type shader, unsigned start_slot,
+ unsigned count, const struct pipe_shader_buffer *buffers,
+ unsigned writable_bitmask)
+{
+ struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
+ unsigned i, idx;
+ for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
+ const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;
+
+ util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);
+
+ if (shader == PIPE_SHADER_VERTEX ||
+ shader == PIPE_SHADER_GEOMETRY) {
+ const unsigned size = buffer ? buffer->buffer_size : 0;
+ const ubyte *data = NULL;
+ if (buffer && buffer->buffer)
+ data = (ubyte *) llvmpipe_resource_data(buffer->buffer);
+ if (data)
+ data += buffer->buffer_offset;
+ draw_set_mapped_shader_buffer(llvmpipe->draw, shader,
+ i, data, size);
+ } else if (shader == PIPE_SHADER_COMPUTE) {
+ llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;
+ } else if (shader == PIPE_SHADER_FRAGMENT) {
+ llvmpipe->dirty |= LP_NEW_FS_SSBOS;
+ }
+ }
+}
+
+static void
+llvmpipe_set_shader_images(struct pipe_context *pipe,
+ enum pipe_shader_type shader, unsigned start_slot,
+ unsigned count, const struct pipe_image_view *images)
+{
+ struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
+ unsigned i, idx;
+
+ draw_flush(llvmpipe->draw);
+ for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
+ const struct pipe_image_view *image = images ? &images[idx] : NULL;
+
+ util_copy_image_view(&llvmpipe->images[shader][i], image);
+ }
+
+ llvmpipe->num_images[shader] = start_slot + count;
+ if (shader == PIPE_SHADER_VERTEX ||
+ shader == PIPE_SHADER_GEOMETRY) {
+ draw_set_images(llvmpipe->draw,
+ shader,
+ llvmpipe->images[shader],
+ start_slot + count);
+ } else if (shader == PIPE_SHADER_COMPUTE)
+ llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;
+ else
+ llvmpipe->dirty |= LP_NEW_FS_IMAGES;
+}
/**
* Return the blend factor equivalent to a destination alpha of one.
*/
-static INLINE unsigned
-force_dst_alpha_one(unsigned factor)
+static inline unsigned
+force_dst_alpha_one(unsigned factor, boolean clamped_zero)
{
switch(factor) {
case PIPE_BLENDFACTOR_DST_ALPHA:
case PIPE_BLENDFACTOR_INV_DST_ALPHA:
return PIPE_BLENDFACTOR_ZERO;
case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
- return PIPE_BLENDFACTOR_ZERO;
+ if (clamped_zero)
+ return PIPE_BLENDFACTOR_ZERO;
+ else
+ return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
}
return factor;
* TODO: there is actually no reason to tie this to context state -- the
* generated code could be cached globally in the screen.
*/
-static void
+static struct lp_fragment_shader_variant_key *
make_variant_key(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
- struct lp_fragment_shader_variant_key *key)
+ char *store)
{
unsigned i;
+ struct lp_fragment_shader_variant_key *key;
- memset(key, 0, shader->variant_key_size);
+ key = (struct lp_fragment_shader_variant_key *)store;
+
+ memset(key, 0, offsetof(struct lp_fragment_shader_variant_key, samplers[1]));
if (lp->framebuffer.zsbuf) {
enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
key->zsbuf_format = zsbuf_format;
memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
}
+ if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
+ key->resource_1d = TRUE;
+ }
+ }
+
+ /*
+ * Propagate the depth clamp setting from the rasterizer state.
+ * depth_clip == 0 implies depth clamping is enabled.
+ *
+ * When clip_halfz is enabled, then always clamp the depth values.
+ *
+ * XXX: This is incorrect for GL, but correct for d3d10 (depth
+ * clamp is always active in d3d10, regardless if depth clip is
+ * enabled or not).
+ * (GL has an always-on [0,1] clamp on fs depth output instead
+ * to ensure the depth values stay in range. Doesn't look like
+ * we do that, though...)
+ */
+ if (lp->rasterizer->clip_halfz) {
+ key->depth_clamp = 1;
+ } else {
+ key->depth_clamp = (lp->rasterizer->depth_clip_near == 0) ? 1 : 0;
}
/* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
if (!lp->framebuffer.nr_cbufs ||
+ !lp->framebuffer.cbufs[0] ||
!util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
key->alpha.enabled = lp->depth_stencil->alpha.enabled;
}
/* alpha.ref_value is passed in jit_context */
key->flatshade = lp->rasterizer->flatshade;
- if (lp->active_occlusion_query) {
+ if (lp->active_occlusion_queries) {
key->occlusion_count = TRUE;
}
}
for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
- enum pipe_format format = lp->framebuffer.cbufs[i]->format;
struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
- const struct util_format_description *format_desc;
- key->cbuf_format[i] = format;
+ if (lp->framebuffer.cbufs[i]) {
+ enum pipe_format format = lp->framebuffer.cbufs[i]->format;
+ const struct util_format_description *format_desc;
- format_desc = util_format_description(format);
- assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
- format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
+ key->cbuf_format[i] = format;
- /*
- * Mask out color channels not present in the color buffer.
- */
- blend_rt->colormask &= util_format_colormask(format_desc);
+ /*
+ * Figure out if this is a 1d resource. Note that OpenGL allows crazy
+ * mixing of 2d textures with height 1 and 1d textures, so make sure
+ * we pick 1d if any cbuf or zsbuf is 1d.
+ */
+ if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
+ key->resource_1d = TRUE;
+ }
- /*
- * Disable blend for integer formats.
- */
- if (util_format_is_pure_integer(format)) {
- blend_rt->blend_enable = 0;
- }
+ format_desc = util_format_description(format);
+ assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
- /*
- * Our swizzled render tiles always have an alpha channel, but the linear
- * render target format often does not, so force here the dst alpha to be
- * one.
- *
- * This is not a mere optimization. Wrong results will be produced if the
- * dst alpha is used, the dst format does not have alpha, and the previous
- * rendering was not flushed from the swizzled to linear buffer. For
- * example, NonPowTwo DCT.
- *
- * TODO: This should be generalized to all channels for better
- * performance, but only alpha causes correctness issues.
- *
- * Also, force rgb/alpha func/factors match, to make AoS blending easier.
- */
- if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
- format_desc->swizzle[3] == format_desc->swizzle[0]) {
- blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
- blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
- blend_rt->alpha_func = blend_rt->rgb_func;
- blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
- blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
+ /*
+ * Mask out color channels not present in the color buffer.
+ */
+ blend_rt->colormask &= util_format_colormask(format_desc);
+
+ /*
+ * Disable blend for integer formats.
+ */
+ if (util_format_is_pure_integer(format)) {
+ blend_rt->blend_enable = 0;
+ }
+
+ /*
+ * Our swizzled render tiles always have an alpha channel, but the
+ * linear render target format often does not, so force here the dst
+ * alpha to be one.
+ *
+ * This is not a mere optimization. Wrong results will be produced if
+ * the dst alpha is used, the dst format does not have alpha, and the
+ * previous rendering was not flushed from the swizzled to linear
+ * buffer. For example, NonPowTwo DCT.
+ *
+ * TODO: This should be generalized to all channels for better
+ * performance, but only alpha causes correctness issues.
+ *
+ * Also, force rgb/alpha func/factors match, to make AoS blending
+ * easier.
+ */
+ if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
+ format_desc->swizzle[3] == format_desc->swizzle[0]) {
+ /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
+ boolean clamped_zero = !util_format_is_float(format) &&
+ !util_format_is_snorm(format);
+ blend_rt->rgb_src_factor =
+ force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
+ blend_rt->rgb_dst_factor =
+ force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
+ blend_rt->alpha_func = blend_rt->rgb_func;
+ blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
+ blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
+ }
+ }
+ else {
+ /* no color buffer for this fragment output */
+ key->cbuf_format[i] = PIPE_FORMAT_NONE;
+ blend_rt->colormask = 0x0;
+ blend_rt->blend_enable = 0;
}
}
*/
key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
+ struct lp_sampler_static_state *fs_sampler;
+
+ fs_sampler = key->samplers;
+
+ memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);
+
for(i = 0; i < key->nr_samplers; ++i) {
if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
- lp_sampler_static_sampler_state(&key->state[i].sampler_state,
+ lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,
lp->samplers[PIPE_SHADER_FRAGMENT][i]);
}
}
if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
for(i = 0; i < key->nr_sampler_views; ++i) {
- if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
- lp_sampler_static_texture_state(&key->state[i].texture_state,
+ /*
+ * Note sview may exceed what's representable by file_mask.
+ * This will still work, the only downside is that not actually
+ * used views may be included in the shader key.
+ */
+ if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
+ lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
}
}
key->nr_sampler_views = key->nr_samplers;
for(i = 0; i < key->nr_sampler_views; ++i) {
if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
- lp_sampler_static_texture_state(&key->state[i].texture_state,
+ lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
}
}
}
+
+ struct lp_image_static_state *lp_image;
+ lp_image = lp_fs_variant_key_images(key);
+ key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
+ for (i = 0; i < key->nr_images; ++i) {
+ if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
+ lp_sampler_static_texture_state_image(&lp_image[i].image_state,
+ &lp->images[PIPE_SHADER_FRAGMENT][i]);
+ }
+ }
+ return key;
}
llvmpipe_update_fs(struct llvmpipe_context *lp)
{
struct lp_fragment_shader *shader = lp->fs;
- struct lp_fragment_shader_variant_key key;
+ struct lp_fragment_shader_variant_key *key;
struct lp_fragment_shader_variant *variant = NULL;
struct lp_fs_variant_list_item *li;
+ char store[LP_FS_MAX_VARIANT_KEY_SIZE];
- make_variant_key(lp, shader, &key);
+ key = make_variant_key(lp, shader, store);
/* Search the variants for one which matches the key */
li = first_elem(&shader->variants);
while(!at_end(&shader->variants, li)) {
- if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
+ if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
variant = li->base;
break;
}
unsigned i;
unsigned variants_to_cull;
- if (0) {
+ if (LP_DEBUG & DEBUG_FS) {
debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
lp->nr_fs_variants,
lp->nr_fs_instrs,
}
/* First, check if we've exceeded the max number of shader variants.
- * If so, free 25% of them (the least recently used ones).
+ * If so, free 6.25% of them (the least recently used ones).
*/
- variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
+ variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
if (variants_to_cull ||
lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
struct pipe_context *pipe = &lp->pipe;
+ if (gallivm_debug & GALLIVM_DEBUG_PERF) {
+ debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
+ "\t%u instrs,\t%u instrs/variant\n",
+ shader->variants_cached,
+ lp->nr_fs_variants, lp->nr_fs_instrs,
+ lp->nr_fs_instrs / lp->nr_fs_variants);
+ }
+
/*
* XXX: we need to flush the context until we have some sort of
* reference counting in fragment shaders as they may still be binned
* Generate the new variant.
*/
t0 = os_time_get();
- variant = generate_variant(lp, shader, &key);
+ variant = generate_variant(lp, shader, key);
t1 = os_time_get();
dt = t1 - t0;
LP_COUNT_ADD(llvm_compile_time, dt);
LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
- llvmpipe_variant_count++;
-
/* Put the new variant into the list */
if (variant) {
insert_at_head(&shader->variants, &variant->list_item_local);
llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
+
+ llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;
+ llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;
}
+
+
projects / mesa.git / blobdiff