(Patch by Teodora Novkovic <teodora.petrovic@gmail.com>, originally reviewed at https://codereview…

skia/ext/convolver_mips_dspr2.cc

+// Copyright (c) 2013 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include <algorithm>
+#include "skia/ext/convolver.h"
+#include "skia/ext/convolver_mips_dspr2.h"
+#include "third_party/skia/include/core/SkTypes.h"
+
+namespace skia {
+// Convolves horizontally along a single row. The row data is given in
+// |src_data| and continues for the num_values() of the filter.
+template<bool has_alpha>
+void ConvolveHorizontally_mips_dspr2(const unsigned char* src_data,
+                                     const ConvolutionFilter1D& filter,
+                                     unsigned char* out_row) {
+#if SIMD_MIPS_DSPR2
+  int row_to_filter = 0;
+  int num_values = filter.num_values();
+  if (has_alpha) {

[Justin Novosad, 2013/05/23 15:32:03]

Since this branch is outside of the loop, we're not getting any performance
benefit from having has_alpha as a template argument.  There is going to be a
run-time conditional branch on has_alpha for each call to this method. The only
difference is whether the branch happens here or in the caller.  Unless there is
tangible evidence (benchmarks?) that prove this has a measurable impact on
performance, this code should all go into the non-template overload below.  On
the other hand, if I am wrong and there is a case for this template, this
function should be in the unnamed namespace (or static).

[Teodora Novkovic, 2013/05/24 16:02:20]

Done.

+    for (int out_x = 0; out_x < num_values; out_x++) {
+      // Get the filter that determines the current output pixel.
+      int filter_offset, filter_length;
+      const ConvolutionFilter1D::Fixed* filter_values =
+        filter.FilterForValue(out_x, &filter_offset, &filter_length);
+      int filter_x = 0;
+
+      __asm__ __volatile__ (
+        ".set push                                  \n"
+        ".set noreorder                             \n"
+
+        "beqz            %[filter_len], 3f          \n"
+        " sll            $t0, %[filter_offset], 2   \n"
+        "addu            %[rtf], %[src_data], $t0   \n"
+        "mtlo            $0, $ac0                   \n"
+        "mtlo            $0, $ac1                   \n"
+        "mtlo            $0, $ac2                   \n"
+        "mtlo            $0, $ac3                   \n"
+        "srl             $t7, %[filter_len], 2      \n"
+        "beqz            $t7, 2f                    \n"
+        " li             %[fx], 0                   \n"
+
+        "11:                                        \n"
+        "addu            $t4, %[filter_val], %[fx]  \n"
+        "sll             $t5, %[fx], 1              \n"
+        "ulw             $t6, 0($t4)                \n" // t6 = |cur[1]|cur[0]|
+        "ulw             $t8, 4($t4)                \n" // t8 = |cur[3]|cur[2]|
+        "addu            $t0, %[rtf], $t5           \n"
+        "lw              $t1, 0($t0)                \n" // t1 = |a0|b0|g0|r0|
+        "lw              $t2, 4($t0)                \n" // t2 = |a1|b1|g1|r1|
+        "lw              $t3, 8($t0)                \n" // t3 = |a2|b2|g2|r2|
+        "lw              $t4, 12($t0)               \n" // t4 = |a3|b3|g3|r3|
+        "precrq.qb.ph    $t0, $t2, $t1              \n" // t0 = |a1|g1|a0|g0|
+        "precr.qb.ph     $t5, $t2, $t1              \n" // t5 = |b1|r1|b0|r0|
+        "preceu.ph.qbla  $t1, $t0                   \n" // t1 = |0|a1|0|a0|
+        "preceu.ph.qbra  $t2, $t0                   \n" // t2 = |0|g1|0|g0|
+        "preceu.ph.qbla  $t0, $t5                   \n" // t0 = |0|b1|0|b0|
+        "preceu.ph.qbra  $t5, $t5                   \n" // t5 = |0|r1|0|r0|
+        "dpa.w.ph        $ac0, $t1, $t6             \n" // ac0+(cur*a1)+(cur*a0)
+        "dpa.w.ph        $ac1, $t0, $t6             \n" // ac1+(cur*b1)+(cur*b0)
+        "dpa.w.ph        $ac2, $t2, $t6             \n" // ac2+(cur*g1)+(cur*g0)
+        "dpa.w.ph        $ac3, $t5, $t6             \n" // ac3+(cur*r1)+(cur*r0)
+        "precrq.qb.ph    $t0, $t4, $t3              \n" // t0 = |a3|g3|a2|g2|
+        "precr.qb.ph     $t5, $t4, $t3              \n" // t5 = |b3|r3|b2|r2|
+        "preceu.ph.qbla  $t1, $t0                   \n" // t1 = |0|a3|0|a2|
+        "preceu.ph.qbra  $t2, $t0                   \n" // t2 = |0|g3|0|g2|
+        "preceu.ph.qbla  $t0, $t5                   \n" // t0 = |0|b3|0|b2|
+        "preceu.ph.qbra  $t5, $t5                   \n" // t5 = |0|r3|0|r2|
+        "dpa.w.ph        $ac0, $t1, $t8             \n" // ac0+(cur*a3)+(cur*a2)
+        "dpa.w.ph        $ac1, $t0, $t8             \n" // ac1+(cur*b3)+(cur*b2)
+        "dpa.w.ph        $ac2, $t2, $t8             \n" // ac2+(cur*g3)+(cur*g2)
+        "dpa.w.ph        $ac3, $t5, $t8             \n" // ac3+(cur*r3)+(cur*r2)
+        "addiu           $t7, $t7, -1               \n"
+        "bgtz            $t7, 11b                   \n"
+        " addiu          %[fx], %[fx], 8            \n"
+
+        "2:                                         \n"
+        "andi            $t7, %[filter_len], 0x3    \n" // residual
+        "beqz            $t7, 3f                    \n"
+        " nop                                       \n"
+
+        "21:                                        \n"
+        "sll             $t1, %[fx], 1              \n"
+        "addu            $t2, %[filter_val], %[fx]  \n"
+        "addu            $t0, %[rtf], $t1           \n"
+        "lh              $t6, 0($t2)                \n" // t6 = filter_val[fx]
+        "lbu             $t1, 0($t0)                \n" // t1 = row[fx * 4 + 0]
+        "lbu             $t2, 1($t0)                \n" // t2 = row[fx * 4 + 1]
+        "lbu             $t3, 2($t0)                \n" // t3 = row[fx * 4 + 2]
+        "lbu             $t4, 3($t0)                \n" // t4 = row[fx * 4 + 2]
+        "maddu           $ac3, $t6, $t1             \n"
+        "maddu           $ac2, $t6, $t2             \n"
+        "maddu           $ac1, $t6, $t3             \n"
+        "maddu           $ac0, $t6, $t4             \n"
+        "addiu           $t7, $t7, -1               \n"
+        "bgtz            $t7, 21b                   \n"
+        " addiu          %[fx], %[fx], 2            \n"
+
+        "3:                                         \n"
+        "extrv.w         $t0, $ac0, %[kShiftBits]   \n" // a >> kShiftBits
+        "extrv.w         $t1, $ac1, %[kShiftBits]   \n" // b >> kShiftBits
+        "extrv.w         $t2, $ac2, %[kShiftBits]   \n" // g >> kShiftBits
+        "extrv.w         $t3, $ac3, %[kShiftBits]   \n" // r >> kShiftBits
+        "sll             $t5, %[out_x], 2           \n"
+        "repl.ph         $t6, 128                   \n" // t6 = | 128 | 128 |
+        "addu            $t5, %[out_row], $t5       \n"
+        "append          $t2, $t3, 16               \n"
+        "append          $t0, $t1, 16               \n"
+        "subu.ph         $t1, $t0, $t6              \n"
+        "shll_s.ph       $t1, $t1, 8                \n"
+        "shra.ph         $t1, $t1, 8                \n"
+        "addu.ph         $t1, $t1, $t6              \n"
+        "subu.ph         $t3, $t2, $t6              \n"
+        "shll_s.ph       $t3, $t3, 8                \n"
+        "shra.ph         $t3, $t3, 8                \n"
+        "addu.ph         $t3, $t3, $t6              \n"
+        "precr.qb.ph     $t0, $t1, $t3              \n"
+        "usw             $t0, 0($t5)                \n"
+
+        ".set pop                                   \n"
+      : [fx] "+r" (filter_x), [out_x] "+r" (out_x), [out_row] "+r" (out_row),
+        [rtf] "+r" (row_to_filter)
+      : [filter_val] "r" (filter_values), [filter_len] "r" (filter_length),
+        [kShiftBits] "r" (ConvolutionFilter1D::kShiftBits),
+        [filter_offset] "r" (filter_offset), [src_data] "r" (src_data)
+      : "lo", "hi", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo", "$ac3hi",
+        "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8"
+      );
+    }
+  } else {
+    for (int out_x = 0; out_x < num_values; out_x++) {
+      // Get the filter that determines the current output pixel.
+      int filter_offset, filter_length;
+      const ConvolutionFilter1D::Fixed* filter_values =
+        filter.FilterForValue(out_x, &filter_offset, &filter_length);
+      int filter_x = 0;
+      __asm__ __volatile__ (
+        ".set push                                  \n"
+        ".set noreorder                             \n"
+
+        "beqz            %[filter_len], 3f          \n"
+        " sll            $t0, %[filter_offset], 2   \n"
+        "addu            %[rtf], %[src_data], $t0   \n"
+        "mtlo            $0, $ac1                   \n"
+        "mtlo            $0, $ac2                   \n"
+        "mtlo            $0, $ac3                   \n"
+        "srl             $t7, %[filter_len], 2      \n"
+        "beqz            $t7, 2f                    \n"
+        " li             %[fx], 0                   \n"
+
+        "11:                                        \n"
+        "addu            $t4, %[filter_val], %[fx]  \n"
+        "sll             $t5, %[fx], 1              \n"
+        "ulw             $t6, 0($t4)                \n" // t6 = |cur[1]|cur[0]|
+        "ulw             $t8, 4($t4)                \n" // t8 = |cur[3]|cur[2]|
+        "addu            $t0, %[rtf], $t5           \n"
+        "lw              $t1, 0($t0)                \n" // t1 = |a0|b0|g0|r0|
+        "lw              $t2, 4($t0)                \n" // t2 = |a1|b1|g1|r1|
+        "lw              $t3, 8($t0)                \n" // t3 = |a2|b2|g2|r2|
+        "lw              $t4, 12($t0)               \n" // t4 = |a3|b3|g3|r3|
+        "precrq.qb.ph    $t0, $t2, $t1              \n" // t0 = |a1|g1|a0|g0|
+        "precr.qb.ph     $t5, $t2, $t1              \n" // t5 = |b1|r1|b0|r0|
+        "preceu.ph.qbra  $t2, $t0                   \n" // t2 = |0|g1|0|g0|
+        "preceu.ph.qbla  $t0, $t5                   \n" // t0 = |0|b1|0|b0|
+        "preceu.ph.qbra  $t5, $t5                   \n" // t5 = |0|r1|0|r0|
+        "dpa.w.ph        $ac1, $t0, $t6             \n" // ac1+(cur*b1)+(cur*b0)
+        "dpa.w.ph        $ac2, $t2, $t6             \n" // ac2+(cur*g1)+(cur*g0)
+        "dpa.w.ph        $ac3, $t5, $t6             \n" // ac3+(cur*r1)+(cur*r0)
+        "precrq.qb.ph    $t0, $t4, $t3              \n" // t0 = |a3|g3|a2|g2|
+        "precr.qb.ph     $t5, $t4, $t3              \n" // t5 = |b3|r3|b2|r2|
+        "preceu.ph.qbra  $t2, $t0                   \n" // t2 = |0|g3|0|g2|
+        "preceu.ph.qbla  $t0, $t5                   \n" // t0 = |0|b3|0|b2|
+        "preceu.ph.qbra  $t5, $t5                   \n" // t5 = |0|r3|0|r2|
+        "dpa.w.ph        $ac1, $t0, $t8             \n" // ac1+(cur*b3)+(cur*b2)
+        "dpa.w.ph        $ac2, $t2, $t8             \n" // ac2+(cur*g3)+(cur*g2)
+        "dpa.w.ph        $ac3, $t5, $t8             \n" // ac3+(cur*r3)+(cur*r2)
+        "addiu           $t7, $t7, -1               \n"
+        "bgtz            $t7, 11b                   \n"
+        " addiu          %[fx], %[fx], 8            \n"
+
+        "2:                                         \n"
+        "andi            $t7, %[filter_len], 0x3    \n" // residual
+        "beqz            $t7, 3f                    \n"
+        " nop                                       \n"
+
+        "21:                                        \n"
+        "sll             $t1, %[fx], 1              \n"
+        "addu            $t2, %[filter_val], %[fx]  \n"
+        "addu            $t0, %[rtf], $t1           \n"
+        "lh              $t6, 0($t2)                \n" // t6 = filter_val[fx]
+        "lbu             $t1, 0($t0)                \n" // t1 = row[fx * 4 + 0]
+        "lbu             $t2, 1($t0)                \n" // t2 = row[fx * 4 + 1]
+        "lbu             $t3, 2($t0)                \n" // t3 = row[fx * 4 + 2]
+        "maddu           $ac3, $t6, $t1             \n"
+        "maddu           $ac2, $t6, $t2             \n"
+        "maddu           $ac1, $t6, $t3             \n"
+        "addiu           $t7, $t7, -1               \n"
+        "bgtz            $t7, 21b                   \n"
+        " addiu          %[fx], %[fx], 2            \n"
+
+        "3:                                         \n"
+        "extrv.w         $t1, $ac1, %[kShiftBits]   \n" // b >> kShiftBits
+        "extrv.w         $t2, $ac2, %[kShiftBits]   \n" // g >> kShiftBits
+        "extrv.w         $t3, $ac3, %[kShiftBits]   \n" // r >> kShiftBits
+        "repl.ph         $t6, 128                   \n" // t6 = | 128 | 128 |
+        "sll             $t8, %[out_x], 2           \n"
+        "addu            $t8, %[out_row], $t8       \n"
+        "append          $t2, $t3, 16               \n"
+        "andi            $t1, 0xFFFF                \n"
+        "subu.ph         $t5, $t1, $t6              \n"
+        "shll_s.ph       $t5, $t5, 8                \n"
+        "shra.ph         $t5, $t5, 8                \n"
+        "addu.ph         $t5, $t5, $t6              \n"
+        "subu.ph         $t4, $t2, $t6              \n"
+        "shll_s.ph       $t4, $t4, 8                \n"
+        "shra.ph         $t4, $t4, 8                \n"
+        "addu.ph         $t4, $t4, $t6              \n"
+        "precr.qb.ph     $t0, $t5, $t4              \n"
+        "usw             $t0, 0($t8)                \n"
+
+        ".set pop                                   \n"
+      : [fx] "+r" (filter_x), [out_x] "+r" (out_x), [out_row] "+r" (out_row),
+        [rtf] "+r" (row_to_filter)
+      : [filter_val] "r" (filter_values), [filter_len] "r" (filter_length),
+        [kShiftBits] "r" (ConvolutionFilter1D::kShiftBits),
+        [filter_offset] "r" (filter_offset), [src_data] "r" (src_data)
+      : "lo", "hi", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo", "$ac3hi",
+        "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8"
+      );
+    }
+  }
+#endif
+}
+template<bool has_alpha>
+void ConvolveVertically_mips_dspr2(const ConvolutionFilter1D::Fixed* filter_val,
+                                   int filter_length,
+                                   unsigned char* const* source_data_rows,
+                                   int pixel_width,
+                                   unsigned char* out_row) {
+#if SIMD_MIPS_DSPR2
+  // We go through each column in the output and do a vertical convolution,
+  // generating one output pixel each time.
+  int byte_offset;
+  int cnt;
+  int filter_y;
+  if (has_alpha) {

[Justin Novosad, 2013/05/23 15:32:03]

Same here.

[Teodora Novkovic, 2013/05/24 16:02:20]

Done.

+    for (int out_x = 0; out_x < pixel_width; out_x++) {
+      __asm__ __volatile__ (
+        ".set push                                   \n"
+        ".set noreorder                              \n"
+
+        "beqz            %[filter_len], 3f           \n"
+        " sll            %[offset], %[out_x], 2      \n"
+        "mtlo            $0, $ac0                    \n"
+        "mtlo            $0, $ac1                    \n"
+        "mtlo            $0, $ac2                    \n"
+        "mtlo            $0, $ac3                    \n"
+        "srl             %[cnt], %[filter_len], 2    \n"
+        "beqz            %[cnt], 2f                  \n"
+        " li             %[fy], 0                    \n"
+
+        "11:                                         \n"
+        "sll             $t1, %[fy], 1               \n"
+        "addu            $t0, %[src_data_rows], $t1  \n"
+        "lw              $t1, 0($t0)                 \n"
+        "lw              $t2, 4($t0)                 \n"
+        "lw              $t3, 8($t0)                 \n"
+        "lw              $t4, 12($t0)                \n"
+        "addu            $t1, $t1, %[offset]         \n"
+        "addu            $t2, $t2, %[offset]         \n"
+        "addu            $t3, $t3, %[offset]         \n"
+        "addu            $t4, $t4, %[offset]         \n"
+        "lw              $t1, 0($t1)                 \n" // t1 = |a0|b0|g0|r0|
+        "lw              $t2, 0($t2)                 \n" // t2 = |a1|b1|g1|r1|
+        "lw              $t3, 0($t3)                 \n" // t3 = |a0|b0|g0|r0|
+        "lw              $t4, 0($t4)                 \n" // t4 = |a1|b1|g1|r1|
+        "precrq.qb.ph    $t5, $t2, $t1               \n" // t5 = |a1|g1|a0|g0|
+        "precr.qb.ph     $t6, $t2, $t1               \n" // t6 = |b1|r1|b0|r0|
+        "preceu.ph.qbla  $t0, $t5                    \n" // t0 = |0|a1|0|a0|
+        "preceu.ph.qbra  $t1, $t5                    \n" // t1 = |0|g1|0|g0|
+        "preceu.ph.qbla  $t2, $t6                    \n" // t2 = |0|b1|0|b0|
+        "preceu.ph.qbra  $t5, $t6                    \n" // t5 = |0|r1|0|r0|
+        "addu            $t6, %[filter_val], %[fy]   \n"
+        "ulw             $t7, 0($t6)                 \n" // t7 = |cur_1|cur_0|
+        "ulw             $t6, 4($t6)                 \n" // t6 = |cur_3|cur_2|
+        "dpa.w.ph        $ac0, $t5, $t7              \n" // (cur*r1)+(cur*r0)
+        "dpa.w.ph        $ac1, $t1, $t7              \n" // (cur*g1)+(cur*g0)
+        "dpa.w.ph        $ac2, $t2, $t7              \n" // (cur*b1)+(cur*b0)
+        "dpa.w.ph        $ac3, $t0, $t7              \n" // (cur*a1)+(cur*a0)
+        "precrq.qb.ph    $t5, $t4, $t3               \n" // t5 = |a3|g3|a2|g2|
+        "precr.qb.ph     $t7, $t4, $t3               \n" // t7 = |b3|r3|b2|r2|
+        "preceu.ph.qbla  $t0, $t5                    \n" // t0 = |0|a3|0|a2|
+        "preceu.ph.qbra  $t1, $t5                    \n" // t1 = |0|g3|0|g2|
+        "preceu.ph.qbla  $t2, $t7                    \n" // t2 = |0|b3|0|b2|
+        "preceu.ph.qbra  $t5, $t7                    \n" // t5 = |0|r3|0|r2|
+        "dpa.w.ph        $ac0, $t5, $t6              \n" // (cur*r3)+(cur*r2)
+        "dpa.w.ph        $ac1, $t1, $t6              \n" // (cur*g3)+(cur*g2)
+        "dpa.w.ph        $ac2, $t2, $t6              \n" // (cur*b3)+(cur*b2)
+        "dpa.w.ph        $ac3, $t0, $t6              \n" // (cur*a3)+(cur*a2)
+        "addiu           %[cnt], %[cnt], -1          \n"
+        "bgtz            %[cnt], 11b                 \n"
+        " addiu          %[fy], %[fy], 8             \n"
+
+        "2:                                          \n"
+        "andi            %[cnt], %[filter_len], 0x3  \n" // residual
+        "beqz            %[cnt], 3f                  \n"
+        " nop                                        \n"
+
+        "21:                                         \n"
+        "addu            $t0, %[filter_val], %[fy]   \n"
+        "lh              $t4, 0($t0)                 \n" // t4=filter_val[fx]
+        "sll             $t1, %[fy], 1               \n"
+        "addu            $t0, %[src_data_rows], $t1  \n"
+        "lw              $t1, 0($t0)                 \n"
+        "addu            $t0, $t1, %[offset]         \n"
+        "lbu             $t1, 0($t0)                 \n" // t1 = row[fx*4 + 0]
+        "lbu             $t2, 1($t0)                 \n" // t2 = row[fx*4 + 1]
+        "lbu             $t3, 2($t0)                 \n" // t3 = row[fx*4 + 2]
+        "lbu             $t0, 3($t0)                 \n" // t4 = row[fx*4 + 2]
+        "maddu           $ac0, $t4, $t1              \n"
+        "maddu           $ac1, $t4, $t2              \n"
+        "maddu           $ac2, $t4, $t3              \n"
+        "maddu           $ac3, $t4, $t0              \n"
+        "addiu           %[cnt], %[cnt], -1          \n"
+        "bgtz            %[cnt], 21b                 \n"
+        " addiu          %[fy], %[fy], 2             \n"
+
+        "3:                                          \n"
+        "extrv.w         $t3, $ac0, %[kShiftBits]    \n" // a >> kShiftBits
+        "extrv.w         $t2, $ac1, %[kShiftBits]    \n" // b >> kShiftBits
+        "extrv.w         $t1, $ac2, %[kShiftBits]    \n" // g >> kShiftBits
+        "extrv.w         $t0, $ac3, %[kShiftBits]    \n" // r >> kShiftBits
+        "repl.ph         $t4, 128                    \n" // t4 = | 128 | 128 |
+        "addu            $t5, %[out_row], %[offset]  \n"
+        "append          $t2, $t3, 16                \n" // t2 = |0|g|0|r|
+        "append          $t0, $t1, 16                \n" // t0 = |0|a|0|b|
+        "subu.ph         $t1, $t0, $t4               \n"
+        "shll_s.ph       $t1, $t1, 8                 \n"
+        "shra.ph         $t1, $t1, 8                 \n"
+        "addu.ph         $t1, $t1, $t4               \n" // Clamp(a)|Clamp(b)
+        "subu.ph         $t2, $t2, $t4               \n"
+        "shll_s.ph       $t2, $t2, 8                 \n"
+        "shra.ph         $t2, $t2, 8                 \n"
+        "addu.ph         $t2, $t2, $t4               \n" // Clamp(g)|Clamp(r)
+        "andi            $t3, $t1, 0xFF              \n" // t3 = ClampTo8(b)
+        "cmp.lt.ph       $t3, $t2                    \n" // cmp b, g, r
+        "pick.ph         $t0, $t2, $t3               \n"
+        "andi            $t3, $t0, 0xFF              \n"
+        "srl             $t4, $t0, 16                \n"
+        "cmp.lt.ph       $t3, $t4                    \n"
+        "pick.ph         $t0, $t4, $t3               \n" // t0 = max_color_ch
+        "srl             $t3, $t1, 16                \n" // t1 = ClampTo8(a)
+        "cmp.lt.ph       $t3, $t0                    \n"
+        "pick.ph         $t0, $t0, $t3               \n"
+        "ins             $t1, $t0, 16, 8             \n"
+        "precr.qb.ph     $t0, $t1, $t2               \n" // t0 = |a|b|g|r|
+        "usw             $t0, 0($t5)                 \n"
+
+        ".set pop                                    \n"
+      : [filter_val] "+r" (filter_val), [filter_len] "+r" (filter_length),
+        [offset] "+r" (byte_offset), [fy] "+r" (filter_y), [cnt] "+r" (cnt),
+        [out_x] "+r" (out_x), [pixel_width] "+r" (pixel_width)
+      : [src_data_rows] "r" (source_data_rows), [out_row] "r" (out_row),
+        [kShiftBits] "r" (ConvolutionFilter1D::kShiftBits)
+      : "lo", "hi", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo", "$ac3hi",
+        "t0", "t1", "t2", "t3", "t4", "t5", "t6","t7", "memory"
+      );
+    }
+  } else {
+    for (int out_x = 0; out_x < pixel_width; out_x++) {
+      __asm__ __volatile__ (
+        ".set push                                   \n"
+        ".set noreorder                              \n"
+
+        "beqz            %[filter_len], 3f           \n"
+        " sll            %[offset], %[out_x], 2      \n"
+        "mtlo            $0, $ac0                    \n"
+        "mtlo            $0, $ac1                    \n"
+        "mtlo            $0, $ac2                    \n"
+        "srl             %[cnt], %[filter_len], 2    \n"
+        "beqz            %[cnt], 2f                  \n"
+        " li             %[fy], 0                    \n"
+
+        "11:                                         \n"
+        "sll             $t1, %[fy], 1               \n"
+        "addu            $t0, %[src_data_rows], $t1  \n"
+        "lw              $t1, 0($t0)                 \n"
+        "lw              $t2, 4($t0)                 \n"
+        "lw              $t3, 8($t0)                 \n"
+        "lw              $t4, 12($t0)                \n"
+        "addu            $t1, $t1, %[offset]         \n"
+        "addu            $t2, $t2, %[offset]         \n"
+        "addu            $t3, $t3, %[offset]         \n"
+        "addu            $t4, $t4, %[offset]         \n"
+        "lw              $t1, 0($t1)                 \n" // t1 = |a0|b0|g0|r0|
+        "lw              $t2, 0($t2)                 \n" // t2 = |a1|b1|g1|r1|
+        "lw              $t3, 0($t3)                 \n" // t3 = |a0|b0|g0|r0|
+        "lw              $t4, 0($t4)                 \n" // t4 = |a1|b1|g1|r1|
+        "precrq.qb.ph    $t5, $t2, $t1               \n" // t5 = |a1|g1|a0|g0|
+        "precr.qb.ph     $t6, $t2, $t1               \n" // t6 = |b1|r1|b0|r0|
+        "preceu.ph.qbra  $t1, $t5                    \n" // t1 = |0|g1|0|g0|
+        "preceu.ph.qbla  $t2, $t6                    \n" // t2 = |0|b1|0|b0|
+        "preceu.ph.qbra  $t5, $t6                    \n" // t5 = |0|r1|0|r0|
+        "addu            $t6, %[filter_val], %[fy]   \n"
+        "ulw             $t0, 0($t6)                 \n" // t0 = |cur_1|cur_0|
+        "ulw             $t6, 4($t6)                 \n" // t6 = |cur_1|cur_0|
+        "dpa.w.ph        $ac0, $t5, $t0              \n" // (cur*r1)+(cur*r0)
+        "dpa.w.ph        $ac1, $t1, $t0              \n" // (cur*g1)+(cur*g0)
+        "dpa.w.ph        $ac2, $t2, $t0              \n" // (cur*b1)+(cur*b0)
+        "precrq.qb.ph    $t5, $t4, $t3               \n" // t5 = |a3|g3|a2|g2|
+        "precr.qb.ph     $t0, $t4, $t3               \n" // t0 = |b3|r3|b2|r2|
+        "preceu.ph.qbra  $t1, $t5                    \n" // t1 = |0|g3|0|g2|
+        "preceu.ph.qbla  $t2, $t0                    \n" // t2 = |0|b3|0|b2|
+        "preceu.ph.qbra  $t5, $t0                    \n" // t5 = |0|r3|0|r2|
+        "dpa.w.ph        $ac0, $t5, $t6              \n" // (cur*r1)+(cur*r0)
+        "dpa.w.ph        $ac1, $t1, $t6              \n" // (cur*g1)+(cur*g0)
+        "dpa.w.ph        $ac2, $t2, $t6              \n" // (cur*b1)+(cur*b0)
+        "addiu           %[cnt], %[cnt], -1          \n"
+        "bgtz            %[cnt], 11b                 \n"
+        " addiu          %[fy], %[fy], 8             \n"
+
+        "2:                                          \n"
+        "andi            %[cnt], %[filter_len], 0x3  \n" // residual
+        "beqz            %[cnt], 3f                  \n"
+        " nop                                        \n"
+
+        "21:                                         \n"
+        "addu            $t0, %[filter_val], %[fy]   \n"
+        "lh              $t4, 0($t0)                 \n" // filter_val[fx]
+        "sll             $t1, %[fy], 1               \n"
+        "addu            $t0, %[src_data_rows], $t1  \n"
+        "lw              $t1, 0($t0)                 \n"
+        "addu            $t0, $t1, %[offset]         \n"
+        "lbu             $t1, 0($t0)                 \n" // t1 = row[fx*4 + 0]
+        "lbu             $t2, 1($t0)                 \n" // t2 = row[fx*4 + 1]
+        "lbu             $t3, 2($t0)                 \n" // t3 = row[fx*4 + 2]
+        "maddu           $ac0, $t4, $t1              \n"
+        "maddu           $ac1, $t4, $t2              \n"
+        "maddu           $ac2, $t4, $t3              \n"
+        "addiu           %[cnt], %[cnt], -1          \n"
+        "bgtz            %[cnt], 21b                 \n"
+        " addiu          %[fy], %[fy], 2             \n"
+
+        "3:                                          \n"
+        "extrv.w         $t3, $ac0, %[kShiftBits]    \n" // r >> kShiftBits
+        "extrv.w         $t2, $ac1, %[kShiftBits]    \n" // g >> kShiftBits
+        "extrv.w         $t1, $ac2, %[kShiftBits]    \n" // b >> kShiftBits
+        "repl.ph         $t6, 128                    \n" // t6 = | 128 | 128 |
+        "addu            $t5, %[out_row], %[offset]  \n"
+        "append          $t2, $t3, 16                \n" // t2 = |0|g|0|r|
+        "andi            $t1, $t1, 0xFFFF            \n"
+        "subu.ph         $t1, $t1, $t6               \n"
+        "shll_s.ph       $t1, $t1, 8                 \n"
+        "shra.ph         $t1, $t1, 8                 \n"
+        "addu.ph         $t1, $t1, $t6               \n" // Clamp(a)|Clamp(b)
+        "subu.ph         $t2, $t2, $t6               \n"
+        "shll_s.ph       $t2, $t2, 8                 \n"
+        "shra.ph         $t2, $t2, 8                 \n"
+        "addu.ph         $t2, $t2, $t6               \n" // Clamp(g)|Clamp(r)
+        "li              $t0, 0xFF                   \n"
+        "ins             $t1, $t0, 16, 8             \n"
+        "precr.qb.ph     $t0, $t1, $t2               \n" // t0 = |a|b|g|r|
+        "usw             $t0, 0($t5)                 \n"
+
+        ".set pop                                    \n"
+      : [filter_val] "+r" (filter_val), [filter_len] "+r" (filter_length),
+        [offset] "+r" (byte_offset), [fy] "+r" (filter_y), [cnt] "+r" (cnt),
+        [out_x] "+r" (out_x), [pixel_width] "+r" (pixel_width)
+      : [src_data_rows] "r" (source_data_rows), [out_row] "r" (out_row),
+        [kShiftBits] "r" (ConvolutionFilter1D::kShiftBits)
+      : "lo", "hi", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo", "$ac3hi",
+        "t0", "t1", "t2", "t3", "t4", "t5", "t6", "memory"
+      );
+    }
+  }
+#endif
+}
+
+void ConvolveVertically_mips_dspr2(const ConvolutionFilter1D::Fixed* filter_val,
+                                   int filter_length,
+                                   unsigned char* const* source_data_rows,
+                                   int pixel_width,
+                                   unsigned char* out_row,
+                                   bool has_alpha) {
+  if (has_alpha) {

[Justin Novosad, 2013/05/23 15:32:03]

This is the conditional branch I was talking about in previous comment. It
executes once per call, so there should be no win from templating, only extra
complexity in the code.

[Teodora Novkovic, 2013/05/24 16:02:20]

Done.

+    ConvolveVertically_mips_dspr2<true>(filter_val,
+                                        filter_length,
+                                        source_data_rows,
+                                        pixel_width,
+                                        out_row);
+  } else {
+    ConvolveVertically_mips_dspr2<false>(filter_val,
+                                         filter_length,
+                                         source_data_rows,
+                                         pixel_width,
+                                         out_row);
+  }
+}
+
+void ConvolveHorizontally_mips_dspr2(const unsigned char* src_data,
+                                     const ConvolutionFilter1D& filter,
+                                     unsigned char* out_row,
+                                     bool has_alpha) {
+  if (has_alpha) {
+    ConvolveHorizontally_mips_dspr2<true>(src_data,
+                                          filter,
+                                          out_row);
+  } else {
+    ConvolveHorizontally_mips_dspr2<false>(src_data,
+                                           filter,
+                                           out_row);
+  }
+}
+} // namespace skia