summaryrefslogtreecommitdiffstats
path: root/src/amd/common
diff options
context:
space:
mode:
authorMarek Olšák <[email protected]>2018-03-20 19:14:57 -0400
committerMarek Olšák <[email protected]>2018-04-02 13:55:22 -0400
commitdc04e4bba2cb3a2831a4413e8de51497cfbd2bc9 (patch)
treeca6b2c96ad9fc912f8e4e430f43dd39e536ae201 /src/amd/common
parenteb779612922e97330a09b9479437f464dd9c6f3b (diff)
radeonsi: move FMASK shader logic to shared code
We'll need it for FBFETCH in both TGSI and NIR paths. Tested-by: Dieter Nützel <[email protected]>
Diffstat (limited to 'src/amd/common')
-rw-r--r--src/amd/common/ac_llvm_build.c56
-rw-r--r--src/amd/common/ac_llvm_build.h3
2 files changed, 59 insertions, 0 deletions
diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 6f577cdac7f..32d8a02f562 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -2435,3 +2435,59 @@ LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
}
return value;
}
+
+/* Adjust the sample index according to FMASK.
+ *
+ * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
+ * which is the identity mapping. Each nibble says which physical sample
+ * should be fetched to get that sample.
+ *
+ * For example, 0x11111100 means there are only 2 samples stored and
+ * the second sample covers 3/4 of the pixel. When reading samples 0
+ * and 1, return physical sample 0 (determined by the first two 0s
+ * in FMASK), otherwise return physical sample 1.
+ *
+ * The sample index should be adjusted as follows:
+ * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
+ */
+void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
+ LLVMValueRef *addr, bool is_array_tex)
+{
+ struct ac_image_args fmask_load = {};
+ fmask_load.opcode = ac_image_load;
+ fmask_load.resource = fmask;
+ fmask_load.dmask = 0xf;
+ fmask_load.da = is_array_tex;
+
+ LLVMValueRef fmask_addr[4];
+ memcpy(fmask_addr, addr, sizeof(fmask_addr[0]) * 3);
+ fmask_addr[3] = LLVMGetUndef(ac->i32);
+
+ fmask_load.addr = ac_build_gather_values(ac, fmask_addr,
+ is_array_tex ? 4 : 2);
+
+ LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
+ fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
+ ac->i32_0, "");
+
+ /* Apply the formula. */
+ unsigned sample_chan = is_array_tex ? 3 : 2;
+ LLVMValueRef final_sample;
+ final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
+ LLVMConstInt(ac->i32, 4, 0), "");
+ final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
+ final_sample = LLVMBuildAnd(ac->builder, final_sample,
+ LLVMConstInt(ac->i32, 0xF, 0), "");
+
+ /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
+ * resource descriptor is 0 (invalid),
+ */
+ LLVMValueRef tmp;
+ tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
+ tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
+ tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
+
+ /* Replace the MSAA sample index. */
+ addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
+ addr[sample_chan], "");
+}
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index f901f336857..8b35028a314 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -414,6 +414,9 @@ LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
unsigned rshift, unsigned bitwidth);
+void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
+ LLVMValueRef *addr, bool is_array_tex);
+
#ifdef __cplusplus
}
#endif